Application of Machine Learning to Epileptic

Seizure Onset Detection and Treatment

by
Ali Hossam Shoeb
Bachelor of Science in Electrical Engineering and Computer Science,
Massachusetts Institute of Technology (2003)
Master of Engineering in Electrical Engineering and Computer
Science, Massachusetts Institute of Technology (2003)
Submitted to the Harvard-MIT Division of Health Sciences and
Technology
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Electrical and Medical Engineering
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2009
@ Massachusetts Institute of Technology 2009. All rights reserved.

Author ...... . ......................

Harvard-MIT Division of Health Sciences and Technology
-September 10, 2009
Certified by. .. ..... . .. .. ..... ...............
C/ John V. Guttag
Professor of Electrical Engineering and Computer Science
r-- . Thesis Supervisor

Accepted by. .... . .............

................... .......
Ram Sasisekharan
Director of Harvard-MIT Division of Health Sciences and Technology
MASS NSl
OF TECHNOLOGY
G4#000ES
OCT 0 2 2009
LIBRARIES
Application of Machine Learning to Epileptic Seizure Onset
Detection and Treatment
by
Ali Hossam Shoeb

Submitted to the Harvard-MIT Division of Health Sciences and Technology

on September 10, 2009, in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy in Electrical and Medical Engineering

Abstract
Epilepsy is a chronic disorder of the central nervous system that predisposes individ-
uals to experiencing recurrent seizures. It affects 3 million Americans and 50 million
people world-wide.
A seizure is a transient aberration in the brain's electrical activity that produces
disruptive physical symptoms such as a lapse in attention and memory, a sensory hal-
lucination, or a whole-body convulsion. Approximately 1 out of every 3 individuals
with epilepsy continues to experience frequent seizures despite treatment with multi-
ple anti-epileptic drugs. These intractable seizures pose a serious risk of injury, limit
the independence and mobility of an individual, and result in both social isolation
and economic hardship.
This thesis presents novel technology intended to ease the burden of intractable
seizures. At its heart is a method for computerized detection of seizure onset. The
method uses machine learning to construct patient-specific classifiers that are capable
of rapid, sensitive, and specific detection of seizure onset. The algorithm detects the
onset of a seizure through analysis of the brain's electrical activity alone or in concert
with other physiologic signals. When trained on 2 or more seizures and tested on
844 hours of continuous scalp EEG from 23 pediatric epilepsy patients, our algorithm
detected 96% of 163 test seizures with a median detection delay of 3 seconds and a
median false detection rate of 2 false detections per 24 hour period.
In this thesis we also discuss how our detector can be embedded within a low-
power, implantable medical device to enable the delivery of just-in-time therapy that
has the potential to either eliminate or attenuate the clinical symptoms associated
with seizures.
Finally, we report on the in-hospital use of our detector to enable delay-sensitive
therapeutic and diagnostic applications. We demonstrate the feasibility of using the
algorithm to control the Vagus Nerve Stimulator (an implantable neurostimulator
for the treatment of intractable seizures), and to initiate ictal SPECT (a functional
neuroimaging modality useful for localizing the cerebral site of origin of a seizure).
Thesis Supervisor: John V. Guttag
Title: Professor of Electrical Engineering and Computer Science
Acknowledgments

I am grateful to the individuals that participated in the clinical studies described in

this thesis. I would like to thank them for their patience, generosity, and courage. I
would like to thank them for teaching me about living with epilepsy, for redefining
my understanding of strength, and for encouraging me to continue my research.
Thanks to Drs. Blaise Bourgeois and S. Ted Treves for their mentorship while
I conducted ictal SPECT studies at Children's Hospital Boston. Thanks to Jack
Connolly and Herman Edwards for teaching me much of what I know about EEG.

Thanks to Dr. Steven Schachter for teaching me how to design clinical studies,
and for exposing me to the clinical management of epilepsy. Thanks to Dr. Steven
Schachter, Dr. Trudy Pang, and the nurses of the Beth Israel Deaconess Medical
Center for making the vagus nerve stimulation studies possible. The prototype system
used in these studies could not have been built without the help of Wayne Ryan and
Ronald Wiken, and the input of professors Steven Leeb, Jeff Lang, Markus Zahn, and
David Perreault.

Thanks to my thesis committee chair, Prof. Tomas Lozano-Perez, for technical

insights that improved the performance of the seizure detection algorithm. Thanks
also to my thesis reader, Dr. Sydney Cash, for introducing me to the fascinating
world of intracranial EEG.

I am grateful to my thesis advisor Prof. John Guttag. John has helped me grow
as an engineer, researcher, teacher, and person over the course of my undergraduate
and graduate careers at MIT.
Thanks to Dorothy Curtis for helping me surpass the hurdles that often get in
the way of a graduate student. Thanks for entertaining endless discussion of drug
pumps, sewing machines, 9V batteries, and electromagnets.
Thanks to Zeeshan Syed and Asfandyar Qureshi for being there throughout both
my undergraduate and graduate years at MIT. Thanks to Eugene Shih for always
encouraging me to do better, and for finding ways to make long days in the lab fun.
Thanks to Jenna Wiens and Anima Singh for cheering me on during the last days
of thesis writing, and for saying "you are going to be here forever". Thanks to Alaa
Kharbouch, Zahi Karam, Georges Aoude, and Ram Srinivasan for always checking
up on me, and for things too many to list. Thanks to James Geraci for being the
best TA. Thanks to Thomas Heldt for being an inspiration. Thanks to Naveen Verma
for an exciting technical collaboration that taught me so much. Thanks to Deema
Arafah for her friendship, and for a memorable and comical discussion about how not
to introduce this work.
Marwa, Fawkia, and Hossam, thanks for always believing in me.
Contents

1 Introduction 27

1.1 Epilepsy . . . . . . . . . . . . . . . . . . . . . . ... ... 27

1.2 Seizure Detection Algorithms and Applications . . . . . . . 29

1.2.1 Applications of Seizure Onset Detection . . . . . . 30

1.2.2 Applications of Seizure Event Detection . . . . . . 30

1.2.3 Application Dependant Performance . . . . . . . . 31

. . . .
1.3 Why is Seizure Detection Challenging? . . . . . . . 31

1.4 Related Work ................... . ... .. 32

1.4.1 Scalp EEG Seizure Event Detectors . . . . . . . . . 33

1.4.2 Scalp EEG Seizure Onset Detectors . . . . . . . . . 34

1.5 Thesis Contributions ............... .. ... . 35

1.6 Thesis Outline ................... .. ... . 38

2 Seizures and the Electroencephalogram

2.1 Epileptic Seizures ...........................

2.2 Scalp Electroencephalogram . ....................

2.2.1 Seizures Within the Scalp Electroencephalogram ......

2.3 Intracranial Electroencephalogram ..................

2.3.1 Seizures Within the Intracranial Electroencephalogram ..

2.4 Summary ...............................

3 Patient-Specific Seizure Onset Detection using the Scalp EEG

3.1 Overview of Binary Classification . .................

 3.2 Feature Vector Design . . . . . . . 56
3.2.1 Spectral Features ......................... 56
3.2.2 Spatial Features .......................... 60
3.2.3 Time Evolution .......................... 63
3.3 Feature Vector Classification . . . . . . . . . . . . . . . . . . . . . . . 66
3.4 Patient-Specific Detector Architecture . . . . . . . . . . . . . . . . . . 70
3.5 Instantiating Detector Parameters . .................. . 73
3.5.1 EEG Epoch Length: L ...................... 73
3.5.2 Number of EEG Channels: N .................. 74
3.5.3 Number of filters: M ....................... 74
3.5.4 Number of feature vectors in Xr: W . . . . . . . . . . . . . . 74
3.5.5 SVM Parameters: -y, J, and C . . . . . . . . . . . . . . . . . . 75
3.5.6 Training Parameters: H, K, and S . . . . . . . . .. . . . . . 75

4 Scalp EEG Data and Testing Methodology 77

4.1 Scalp EEG Data Set ............................ 77
4.2 Performance Metrics .......................... . 78
4.3 Performance Metric Measurement . . . . . . . ....... .. . . . . . . 78

5 Performance 81
5.1 Patient-Specific Detector Performance . . . . . . . ... ... . . 81
5.1.1 Latency . . . . . . . .. . . . . . . .. . . . . . . . . . . . . . . 81
5.1.2 Sensitivity ........... ...... .......... . 82
5.1.3 Specificity ............ ... ............... 82
5.2 Varying Patient-Specific Detector Parameters . . . . . . . . .... . 87
5.2.1 Varying the Number of Filters: M . . . . . . . . . . . . . . . 87
5.2.2 Varying the Number of Feature Vectors in Xr: W . . . . . . 88
5.2.3 Varying the Number of Training Seizures: K .......... 89
5.2.4 Varying Training Time into a Seizure: S . ...... .... . 90
5.3 Patient-Specific and Non-specific Seizure Detection .... . .... . . 92
5.3.1 Performance Comparison . . . . . . . . . . . . . . . . . . . . . 92
 5.4 Case Studies . ..........................
5.4.1 Latency .........................
5.4.2 Sensitivity ........................
5.4.3 Specificity ........................

6 Seizure Onset Detection Using Physiologic Signal Fusion 101

6.1 Why Use a Second Physiologic Signal? ............ 101

6.2 The Electrocardiogram as a Second Signal .......... 102

6.3 Patient-Specific, EEG-ECG-based Seizure Detection ..... 102

6.4 Case Studies ... .... ... . ... ... .. ... . .. .. 104

6.4.1 D ata . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.4.2 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.4.3 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 106

6.5 Importance of Patient-Specificity . .............. 111

7 Seizure-Triggered Vagus Nerve Stimulation 113

7.1 Vagus Nerve Stimulation .................... 113

7.2 Methods ......... 115

7.2.1 System Overview 115

7.2.2 Study Protocol . 116

7.3 Case Studies ....... 117
7.3.1 Patient A . . . . 117
7.3.2 Patient B . . . . 121

7.3.3 Patient C . ... : : : : : : : 125

8 Seizure-Triggered Single Photon Emission Computed Tomography 129

8.1 SPECT in Epilepsy ............. . . . . . . . . 129
8.2 M ethods ................... . . . . . . . . 132
8.2.1 System Overview . ......... . . . . . . . . 132
8.2.2 Study Procotol ........... . . . . . . . . 132
8.3 Results of Clinical Evaluation . . . . . . . . . . . . . . . 134
9 Patient-Specific Seizure Onset Detection using iEEG 137
9.1 Why Detect Seizure Onset Using iEEG? ................ 137
9.2 Patient-Specific Detector Architecture ................... 138
9.2.1 Cost of Support-Vector Machine Classification . . . . . . . . . 139
9.3 Patient Non-specific Detector Architecture . . . . . . . . . . . . . . . 142
9.4 iEEG Data Set .............................. 143
9.5 Performance Comparison . . . . . . . . . . . . . . . . . . . . . . . . 143
9.6 Reduced and Non-Reduced Support-Vector Machines . . . . . . . . . 145
9.7 Case Studies . . . . . . . . . . . . . . . . . . . . . . . . ....
. ..... . 148
9.7.1 Latency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
9.7.2 False Detections . . . . . . . . . . . . . . . ............. . 149
9.8 Predicting Clinical Seizure Onset . . . . . . . . . . . . . . . . . . . . 149
9.9 Implementation .............................. 150

10 Conclusion and Future Work 153

10.1 Goals and Contributions ......................... .. 153
10.2 Future Work ........................ ...... . 155
10.2.1 Feature Vector Enhancement . . . . . . . . . . . . . . . . . . 155
10.2.2 Detecting Seizure Cessation . . . . . . . . . . . . . . . . . . . 155
10.2.3 Closed-Loop, Non-invasive Brain Stimulation . . . . . . . . . . 156
List of Figures

2-1 EEG electrodes arrayed symmetrically across the scalp provide a tem-
poral and spatial summary of the synchronous firing of tens of millions
of neurons within the brain. ....................... 41

2-2 Example of 10 seconds of awake EEG interrupted by an eye-blink at 37

seconds. The eye-blink results in a downward deflection on the EEG

channels {FP1 - F7, FP1 - F3, FP2 - F4, FP2 - F8}. ...... . 42

2-3 Example of 10 seconds of sleep EEG interrupted by 11 Hz oscillations

known as sleep spindles. The sleep spindles are most visible on the
channel FP2 - F4 between 8-10 and 12-14 seconds. . .......... 43

2-4 Example of 10 seconds of awake EEG interrupted by the rhythmic,

high-frequency activity associated with chewing. The high-frequency
activity caused by chewing can be seen on the EEG channels F7 - T7
and F8-T8. ........... ................... 43

2-5 Example of a seizure within the scalp EEG of Patient A. The seizure,
which begins at 1723 seconds, involves flattening of the EEG signal
across all channels followed by the appearance of a beta band rhythm
on the channels {F3 - C3, C3 - P3}................... 44

2-6 Example of a seizure within the scalp EEG of Patient B. The seizure,

which begins at 6313 seconds, involves the appearance of a theta band

rhythm on the channels {F7 - T7, T7 - P7}. . ........... . 45

2-7 Example of an abnormal discharge within the scalp EEG of Patient
A. The discharge, which occurs between 2884-2892 seconds, involves
a repeating pattern of high-amplitude spikes followed by broad waves
that can be seen on most EEG channels. . .............. . 46
2-8 Example of abnormal rhythmic activity within the scalp EEG of Pa-
tient B. The rhythmic activity, which occurs between 6126-6130 sec-
onds, involves a theta band rhythm on the frontal channels {FP1 -
F3,F3 - C3} ................ .... ......... ... 47
2-9 A second seizure within the scalp EEG of Patient A. The second
seizure, which begins at 6210 seconds, resembles the seizure shown
in Figure 2-5. ............................... 47
2-10 A second seizure within the scalp EEG of Patient B. The second
seizure, which begins at 2381 seconds, resembles the seizure shown
in Figure 2-6. ............................... 48
2-11 Example of a seizure within the scalp EEG of Patient C. The seizure,
which begins at 1486 seconds, involves rapid eye-blinking followed by
the appearance of a 3-4 Hz theta wave on the channel T4 - T6. . . . 49
2-12 Example of a seizure recorded from Patient C approximately 10 months
after recording the seizure in Figure 2-11. . . . . . . . . . . . . . . . . 49
2-13 Example of seizures within the iEEG of patients D (top panel) and
C (bottom panel). The seizure of Patient D begins at 5 seconds and
consists of a few spikes that evolve into a high-amplitude spike train.
The seizure of Patient E begins at 5 seconds and consists of a few spikes
followed by low-amplitude, high-frequency activity. . ......... 51
2-14 The iEEG of Patient F contains rhythmic activity that is associated
with a seizure (top panel) as well as rhythmic activity that is not
associated with any clinical symptoms (bottom panel). . ........ 52
2-15 Example of two seizures within the iEEG of Patient D. Both seizures
begin at 5 seconds and consist of a few spikes that evolve into a high-
amplitude spike train. .......................... 52
3-1 Example of a seizure whose onset is associated with rhythmic activity
that contains a mixture of prominent frequency components. Seizure

onset follows 2994 seconds and mostly involves channels on the right
side of the head (channels with even numerals). . ............ 57

3-2 Frequency spectrum of the channel FP2- F4 following the onset of the

seizure illustrated in Figure 3-1. The spectrum contains large spectral

components at 2, 5, and 11 Hz. ..................... 57

3-3 Superposition of the frequency spectra of an eye-blink and the rhyth-

mic activity observed on the channel FP2 - F4 following the onset

of the seizure in Figure 3-1. The seizure spectrum contains an 11 Hz
component that is absent from the spectrum of the eye-blink . . .. 58

3-4 Superposition of the frequency spectra of a sleep spindle and the rhyth-
mic activity observed on the channel FP2- F4 following the onset of
the seizure in Figure 3-1. The seizure spectrum contains low-frequency
spectral components that are larger than those in the spindle spectrum. 59

3-5 Superposition of the frequency spectra of chewing and the rhythmic

activity observed on the channel FP2 - F4 following the onset of
the seizure in Figure 3-1. The seizure spectrum contains less high-

frequency content relative to the chewing spectrum. . ......... 59

3-6 M-band filterbank that measures the energy within the spectral com-
ponents of an L second epoch taken from a single EEG channel. . . . 60

3-7 Energy within the frequency bands defined by the rightmost three fil-
ters (thick line) of the filterbank differentiates between the seizure spec-
trum (red) and chewing spectrum (blue) . ............... 61

3-8 The spatial distribution of EEG activity can be used to differentiate

seizure from the non-seizure activity. The sleep spindles between 2989-

2992 seconds do no involve activity on the channels C4 - P4 and T8 -

P8. In contrast, the seizure activity following 2994 seconds involves

both of these channels ................... ....... 61

3-9 Formation of an intermediate feature vector XT that captures the spec-
tral and spatial properties of an epoch at time t = T . ........ 62

3-10 Scalp EEG seizure involving a sequence of discrete spectral events.

The first event, at 1723 seconds, is a spike involving all EEG channels.
Next, all EEG channels exhibit a period of low-amplitude EEG. Finally,
at 1725 seconds, a beta band rhythm that increases in amplitude and
decreases in frequency appears on the channel F3 - C3. . ....... 64

3-11 Spectrogram of Channel F3 - C3 illustrating sequence of events that

compose the onset and evolution of the seizure in Figure 3-10. .... 65

3-12 Spectrogram of channel F3 - C3 illustrating how the feature vector -

models the transition from background to seizure onset. The feature
vector XT captures the period of low amplitude activity, XT-2 captures
the spike, and XT-4 captures the background EEG activity preceding
the spike .................................. 65

3-13 Spectrogram of channel F3- C3 illustrating how the feature vector 2T

models the evolution of a seizure. The feature vector XT captures the
beta band rhythm, XT-2 captures the period of low amplitude activity,
and XT-4 captures the spike. ....................... 66

3-14 Two-dimensional seizure (red) and non-seizure (blue) feature vectors

extracted from a single iEEG channel. The first dimension of this
feature space is defined by the energy in the spectral band 0-16 Hz
and the second dimension corresponds to the energy in the 25 + 11
Hz band. The seizure vectors extracted from the first 7 seconds of the
seizure are numbered 1-5 . .. .............. ...... . 67

3-15 Linear decision boundary separating seizure (red) and non-seizure (blue)
feature vectors extracted from a single iEEG channel. The decision
boundary was determined using the SVM learning algorithm...... 68
3-16 Nonlinear decision boundary separating seizure (red) and non-seizure
(blue) feature vectors extracted from a single iEEG channel. The deci-
sion boundary was determined using a radial basis kernel and the SVM

learning algorithm. ............................ 69

3-17 Patient-specific seizure onset detector architecture. . ........ . 71

3-18 Generating training non-seizure feature vectors. . ........ . 72

3-19 Generating training seizure feature vectors. ........... . . . . 73

5-1 Percentage of 163 test seizures detected within a specified latency. The
detector notes the onset of 50% of all test seizures within 3 seconds. . 82

5-2 Latency with which the detector notes the onset of seizures for each
of the 23 test subjects. Each dot represents a seizure. A numeral is
placed next to dots that represent more than one seizure. For most
patients, the majority of seizures are detected within 5 seconds. . . . 83

5-3 Example of a seizure within the scalp EEG of Patient 15. The seizure,
which begins at 272 seconds, consists of a theta band rhythm that is

most prominently seen on the channel T7 - P7. . .......... . 83

5-4 Example of another type of seizure observed within the scalp EEG of
Patient 15. This seizure, which begins at 876 seconds, consists of a
train of spikes on the channel P7 - 01. The detector fails to detect
the onset of this seizure since its spectral and spatial characteristics
differ from those of training seizures. Training seizures resemble the
seizure shown in Figure 5-3. ....................... 84

5-5 Sensitivity of the patient-specific seizure detector. Red bars show the
number of test seizures available for each subject, and black bars show
the number of test seizures recognized by the detector. Overall, the
detector recognized 96% of 163 test seizures. . .......... . 84

5-6 Specificity of the patient-specific seizure detector. For most patients

(18 of 23), the detector declares between 0 and 5 false detections per
24 hours................................. .. 85
5-7 Example of a seizure within the scalp EEG of Patient 13. The seizure,
which begins at 640 seconds, consists of a delta band rhythm that is
most prominent on the channel FP1 - F7. . ............. . 86

5-8 Example of a short burst of rhythmic activity within the scalp EEG of
Patient 13. The burst, which occurs between 233-235 seconds, consists
of a delta band rhythm that is most prominent on the channel FP1 -
F7. Since the burst resembles the seizure in Figure 5-7, the detector
declares a false detection following the onset of the burst. ....... 86

5-9 False detections declared by the detector are not uniformly distributed
across test records. For Patient 13, 75% of the test records resulted in
0 false detections. For Patient 12, 68% of the test records resulted in
0 false detections .................... . ......... 87

5-10 The false detection rate of the detector decreases as number of filters
used to construct XT is increased from 2 to 8. A two-filter filterbank
yields a false detection rate of 8 false detections per 24 hours. An
eight-filter filterbank results in a false detection rate smaller than 4
false detections per 24 hours ........................ 88

5-11 Increasing the number of feature vectors within X7 decreases the de-
tector's false detection rate and increases its detection delay. To detect
seizures with an average latency shorter than 5 seconds one should set
W = 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . 89

5-12 Increasing the number of training seizures decreases the detector's miss
rate and its detection delay. Including more than three training seizures
results in a marginal improvement in detector's miss rate........ 90

5-13 Including seizure vectors derived from the first S > 12 seconds of each
training seizure does not improve seizure detection delay significantly.
For S > 12 seconds the detector's mean latency is less than 3 seconds. 91
5-14 Including seizure vectors derived from the first S > 18 seconds of each
training seizure does not improve seizure detection rate significantly.
For S > 18 seconds the detector recognizes more than 95% of all test
seizures. .................................. 91

5-15 Increasing S increases the detector's false detection rate. ....... . 92

5-16 Comparison of the number of test seizures recognized by the Reveal

(sensitive setting) and patient-specific algorithms. Black bars show the

number of test seizures available for each subject, red bars show the
number of test seizures recognized by the patient-specific algorithm,
and blue bars show the number of seizures recognized by the Reveal
algorithm. The Reveal algorithm detected 74% of 152 test seizures and

our patient-specific method detected 96% of all test seizures. ..... 93

5-17 Comparison of the number of false detections declared by the Reveal

(sensitive setting) and patient-specific algorithms. For some patients

the Reveal algorithm declared an excess of 100 false detections per 24
hour period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

5-18 Comparison of the number of false detections declared by the Reveal

(sensitive setting) and patient-specific algorithms. Expanded y-axis. . 94

5-19 Comparison of the number of test seizures recognized by the Reveal

algorithm (specific setting) and our patient-specific algorithm. Black
bars show the number of test seizures available for each subject, red
bars show the number of test seizures recognized by the patient-specific

algorithm, and blue bars show the number of seizures recognized by

the Reveal algorithm. The Reveal algorithm detected 61% of 152 test
seizures and our patient-specific method detected 96% of all test seizures. 95

5-20 Comparison of the number of false detections declared by the Reveal

(specific setting) and patient-specific algorithms. For some patients

the Reveal algorithm declared an excess of 100 false detections per 24

hour period .............. .. ................ .. 96

5-21 Comparison of the number of false detections declared by the Reveal
(specific setting) and patient-specific algorithms. Expanded y-axis.. . 96

5-22 Seizure recorded within the scalp EEG of Patient G. The seizure, which
begins at 701 seconds, consists of beta band activity that is most promi-
nent on the channels {F4 - C4, C4 - P4,F8 - T8, T8 - P8}. .... 97

5-23 Large amplitude theta band activity can be seen on all EEG channels
twenty-nine seconds after the onset of the seizure shown in Figure 5-22. 98

5-24 Seizure recorded within the scalp EEG of Patient H. The seizure, which
begins at 977 seconds, consists of rapid eye-blinking that is later ac-
companied by theta band activity on the channels {T4 - T6, T6 - 02}

at 983 seconds ............................... 99

5-25 Seizure recorded within the scalp EEG of Patient J. The onset of this
seizure invovles a generalized spike at 12231 seconds that is followed
by generalized rhythmic activity. ..................... 100

5-26 Rhythmic activity commonly observed within the awake scalp EEG of
Patient J. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

6-1 Block diagram of patient-specific seizure onset detector that combines

features extracted from the EEG and ECG signals. . .......... 103

6-2 Example of a seizure within the scalp EEG of Case 1. The seizure,
which begins at 1486 seconds, involves rapid eye-blinking that results in
downward deflections on frontal EEG channels (e.g. {FP1-F3,FP2-
F4}). Coincident with the onset of rapid eye-blinking, the patient's
heart rate accelerates as shown in Figure 6-3. Later, at 1492 seconds,
a 3-4 Hz theta wave appears on the EEG channel T4 - T6. ....... 105

6-3 Seizure onset, at 1486 seconds, is associated with an acceleration of

the patient's heart rate. ......................... 106
6-4 Comparison of the minimum, maximum, and mean detection delays of
two detectors. One detector classifies a feature vector composed solely
of EEG features, while the other uses a feature vector that combines
EEG and ECG features. The detector that fuses features extracted
from the EEG and ECG signals has a shorter mean detection delay. . 107

6-5 Comparison of the false detection rates of two detectors. One detector
classifies a feature vector composed solely of EEG features, while the
other uses a feature vector that combines EEG and ECG features. The
detector that fuses features extracted from the EEG and ECG signals
has a smaller false detection rate. .................... 107

6-6 Example of a seizure within the scalp EEG of Case 2. The seizure,
which begins at 56 seconds, involves a 12 second period of low-amplitude
EEG activity across most EEG channels. At the same time, the pa-
tient's heart rate accelerates as shown in Figure 6-7. Later, at 68
seconds, 1-2 Hz generalized, rhythmic activity develops. ......... 108

6-7 Seizure onset, at 56 seconds, is associated with an acceleration of the

patient's heart rate ................... ......... 109

6-8 Comparison of the minimum, maximum, and mean detection delays of

two detectors. One detector classifies a feature vector composed solely
of EEG features, while the other uses a feature vector that combines
EEG and ECG features. The two detectors have comparable seizure
detection delays. ............................. 110

6-9 Comparison of the false detection rates of two detectors. One detector
classifies a feature vector composed solely of EEG features, while the
other uses a feature vector that combines EEG and ECG features. The
detector that fuses features extracted from the EEG and ECG signals
has a smaller false detection rate. . .................. . 110

7-1 Block diagram of closed-loop vagus nerve stimulation system. ..... 116
7-2 Example of a seizure within the scalp EEG of Patient A. The seizure,
which begins at 1486 seconds, involves rapid eye-blinking that results in
downward deflections on frontal EEG channels (e.g. {FP1-F3,FP2-
F4}). Coincident with the onset of rapid eye-blinking, the patient's
heart rate accelerates as shown in Figure 7-3. Later, at 1492 seconds,
a 3-4 Hz theta wave appears on the EEG channel T4 - T6. ....... 118

7-3 Seizure onset, at 1486 seconds, is associated with an acceleration of

the patient's heart rate. ............... ........ .. 119

7-4 Block diagram of patient-specific seizure onset detector that combines

features extracted from the EEG and ECG signals. . .......... 120

7-5 Initiation of VNS following computerized detection of the onset of a

seizure from Patient A. Seizure onset begins with rapid eye blinking at
992 seconds. The detector declared seizure onset at 996 seconds, and
initiated VNS in response. Beginning at 1002 seconds, a spike-train
appears on the "VNS" channel confirming the initiation of vagus nerve
stimulation. ...... .. ........ .. ............ . 120

7-6 Seizure onset, at 992 seconds, is associated with an acceleration of the

patient's heart rate. . .................. ......... 121

7-7 Typical EEG change associated with onset of Patient B's seizure. The
seizure, which begins at 1340 seconds, involves an electrodecrement
most prominent on the occipital channels P3 - 01 and P4- 02. Next,
at the 1348 seconds, a 3-5 Hz occipital rhythm emerges from the elec-
trodecrement and rapidly generalizes. By 1360 seconds, muscle activity
associated with the tonic-clonic phase of the seizure becomes visible. . 123

7-8 Block diagram of patient-specific seizure onset detector that uses fea-
tures extracted from the EEG signal. . .................. 123
7-9 Initiation of VNS following computerized detection of the onset of a
seizure from Patient B. The seizure begins with an electrodecrement at
1485 seconds. At 1501 seconds, the computerized system detected the
muscle activity associated with the tonic-clonic phase of the seizure
and initiated VNS in response. Evidence of VNS generator activity
can be seen on the VNS channel at the conclusion of the seizure as
shown in Figure 7-10 ............................ 124

7-10 The spike-train seen on the VNS channel between 1555-1565 seconds
confirms the automatic initiation of vagus nerve stimulation following
the onset of the seizure illustrated in Figure 7-9. . .......... . 124

7-11 Example of a seizure within the scalp EEG of Patient C. The seizure,
which begins at 248 seconds, involves a 14 second period of low-amplitude
EEG activity across most EEG channels. Later, at 262 seconds, gen-
eralized, rhythmic activity develops. . ................. . 126

7-12 Initiation of VNS following computerized detection of the onset of a

seizure from Patient C. The seizure begins with an electrodecrement
at 56 seconds. The computerized system noted seizure activity at 59
seconds and initiated vagus nerve stimulation in response. Evidence of
VNS generator activity can be seen on the VNS channel starting at 69
seconds. . .................. . . ............ .. 127

7-13 The spike-train seen on the VNS channel (first channel from the bot-
tom), between 68-84 seconds, confirms the automatic initiation of va-
gus nerve stimulation following the onset of the seizure illustrated in
Figure 7-12. ................................ 127

8-1 SPECT Image taken outside the seizure state (Interictal) and during
the seizure state (Ictal I). During a seizure the seizure focus is hyper-
perfused and appears as a bright spot in a SPECT Image. The third
image (Ictal II) is the result of subtracting the first two images. . . 130
8-2 Delay associated with injecting the ictal SPECT radiotracer ranges be-
tween 10-130 seconds for 18 pediatric patients undergoing ictal SPECTs

at Children's Hospital Boston. ..................... 131

8-3 Block diagram of system for automatic infusion of ictal SPECT radio-
tracer. ................................... 132

8-4 Protocol for evaluating automatic system for infusion of ictal SPECT
radiotracer. See section 8.2.2 for details. . ........... . . . . 133

8-5 Comparison of delay between seizure onset and completion of radio-

tracer infusion by the clinical team (Group 1) and the computerized
system (Group 2) .............................. 135

8-6 Latency and number of false detections declared by the computerized

system during eight Ictal SPECT studies ................. 136

9-1 Block diagram of a patient-specific seizure detection algorithm that

uses spectral and spatial features extracted from two iEEG channels. 139

9-2 Superposition of a nonlinear SVM boundary requiring Nsv = 50 terms

(solid line) and an approximation of that boundary using Msv = 8

terms (dashed line) ........... ..... ........... 142

9-3 Block diagram of patient non-specific seizure detection algorithm. .. 143

9-4 Comparison of the detection latencies of the patient-specific and pa-

tient non-specific seizure detectors. The patient-specific detector that

used a nonlinear SVM detected seizures within 8.5 ± 5.1 seconds. The

same detector using a linear SVM detected seizures within 9.3±4.8 sec-

onds. The patient non-specific algorithm (with T=3) detected seizures

within 10.5 ± 8.7 seconds. ........................ 144

9-5 Comparison of the number of false detections declared by the patient-
specific and patient non-specific seizure detectors during 40 hours of
non-seizure data. The patient-specific detector with a nonlinear SVM

declared 19 false detections. The same detector using a linear SVM

declared 28 false detections. The patient non-specific detector (with

T=3) declared 126 false detections. . ................... 145

9-6 Comparison of the number of false detections declared by the patient-

specific and patient non-specific seizure detectors during 40 hours of

non-seizure data. The patient-specific detector with a nonlinear SVM

declared 19 false detections. The same detector using a linear SVM
declared 28 false detections. The patient non-specific detector (with
T=10) declared 17 false detections. . ................... 146

9-7 Comparison of the detection latencies of the patient-specific and pa-

tient non-specific seizure detectors. The patient-specific detector that
used a nonlinear SVM detected seizures within 8.5 ± 5.1 seconds. The
same detector using a linear SVM detected seizures within 9.3 ± 4.8
seconds. The patient non-specific algorithm (with T=10) detected

seizures within 18.7 ± 10.8 seconds. . ................... 146

9-8 Comparison of latency of detectors that use support-vector machines

with different number of support-vectors . ............... 147

9-9 Comparison of number of false detections declared by detectors that

use support-vector machines with different number of support-vectors 147

9-10 Example of a seizure within the iEEG of Patient 6. The seizure, which
begins at 2 seconds, consists of a few spikes that evolve into a high-
amplitude spike train. The patient-specific detector recognized the
seizure at 16 seconds, while the patient non-specific detector did so at
40 seconds. ............................... 148

9-11 Example of a seizure within the iEEG of Patient 5. The seizure, which
begins at 6 seconds, consists of a high-frequency rhythm that increases

in amplitude and decreases in frequency as the seizure progresses. . . 149

9-12 Example of a burst of rhythmic non-seizure activity within the iEEG
of Patient 5. The patient non-specific detector declared this burst as
a seizure event even though its spatial and spectral character differs
significantly from that of the seizure shown in Figure 9-11. ....... 150
9-13 Detection delay relative to the clinical onset of a seizure. An alarm
based on detecting the electrical onset of a seizure using our patient-
specific method could provide some patients, such as patients 3 and 6,
enough time to prepare for the clinical onset of a seizure. ........ 151
9-14 Implementation of our machine-learning based, patient-specific detec-
tor on the hardware described in [2]. An analog front-end processes
two iEEG channels, and for each channel, extracts the spectral power
within two configurable frequency bands. A digital back-end samples
the analog power profile, assembles these samples into a feature vector,
and classifies the feature vector using a support-vector machine. . . . 151
List of Tables

3.1 Patient-specific Detector Parameters . ................. 74

26
Chapter 1

Introduction

1.1 Epilepsy

Epilepsy is a chronic disorder of the central nervous system that predisposes individ-
uals to experiencing recurrent seizures. A seizure is a sudden, transient aberration
in the brain's electrical activity that produces disruptive symptoms. These symp-
toms range between a lapse in attention, a sensory hallucination, or a whole-body
convulsion. In the book "Brainstorms: Epilepsy in Our Words" [49], individuals with
epilepsy describe what it is like to have a seizure:

I only experience my seizures as I am falling off to sleep...my symptoms

include a shock-like feeling inside my head, a twitch of one or more limbs,
a shock that makes the trunk of my body jump.

I experience a combination of deja vu with extreme fear. Nothing I do

takes me out of the deja vu...the general feeling is of being in front of an
oncoming train with no way to escape.

Epilepsy is not a single disease, but a family of syndromes that share the feature
of recurrent seizures. Epilepsy may develop as a result of inheriting a mutation in
a molecular mechanism that regulates neuron behavior, migration, or organization.
Alternatively, it may develop as a result of brain trauma such as a severe blow to the
head, a stroke, a cerebral infection, or a brain malignancy [8].
 Fifty million people world-wide are diagnosed with epilepsy. In the United States

epilepsy affects 3 million people and is the third most common neurologic disorder
after Alzheimer's disease and stroke. In an unfortunate subset of 1.2 million indi-
viduals, frequent, unpredictable seizures persist despite treatment by one or multi-

ple anti-epileptic drugs. These types of seizures are known as medically intractable

seizures.

The worst part of having seizures is knowing that they can happen anytime
and even though drugs control mine most of the time they occasionally
break through.

Medically intractable seizures severely limit the independence and mobility of an

individual and, as a consequence, can result in social isolation and economic hardship.
Most concerning is that refractory seizures significantly increase an individual's chance

of experiencing burns, lacerations, skull fractures, and even sudden unexpected death
[16, 17].

I knew, also instantaneously, that I must keep this affliction secret, and
that this condition and the dark secret of it set me apart from others. I
was to live my life feeling this isolation and separation.

The negative influence of uncontrolled seizures extends beyond the individual to

affect their family members, friends, and the whole of society. The families and friends
of people with epilepsy experience chronic anxiety and rearrange their lives to ensure
the safety of their loved one. Society incurs an annual loss of 12.5 billion dollars in
health care costs and losses in productivity [6]. There is a need for novel therapies
that better control seizures as well as technology that helps both the individual and

their family to cope with the consequences of seizures.

Computerized seizure onset detection will enable the engineering of novel ther-

apeutic and alerting systems that may ease the burden of intractable seizures. A

therapeutic system capable of detecting and reacting to the onset of a seizure could
administer a local electrical [60], thermal [47], or neurochemical [69] stimulus that
halts the progression of a seizure prior to the development of clinical symptoms.
Moreover, just-in-time, local therapy could relieve patients of the toxic side-effects
that accompany systemic administration of multiple anti-epileptic drugs. An alerting
system equipped with seizure onset detection could warn the patient of the seizure
prior to the development of debilitating symptoms, or could notify a family member
so that the consequences of a seizure are limited. Knowledge that a reliable warn-
ing will be issued rapidly following seizure onset may restore within individuals the
confidence to overcome the limits on life that accompany seizures.

In this thesis, we describe a patient-specific algorithm capable of rapidly detecting

seizure onset through analysis of the brain's electrical activity alone or in concert
with other physiologic signals. Moreover, we demonstrate the feasibility of using

our detector to control the Vagus Nerve Stimulator (an implantable neurostimulator
for the treatment of intractable seizures), and to initiate ictal SPECT (a functional
neuroimaging modality useful for localizing the cerebral site of origin of a seizure).

In this chapter, we describe different types of seizure detectors and their potential

role in diagnosistic, therapeutic, and alerting applications. We also highlight the

challenges associated with the computerized detection of seizures using the brain's
electrical activity. Finally, we review the methodology and performance of previously
published algorithms and highlight how our approach contributes to the field of seizure
detection.

1.2 Seizure Detection Algorithms and Applications

A seizure detector can be classified as either a seizure onset detector or as a seizure

event detector. The purpose of a seizure onset detector is to recognize that a seizure
has started with the shortest possible delay, but not necessarily with the highest
possible accuracy. In contrast, the purpose of a seizure event detector is to identify

seizures with the greatest possible accuracy, but not necessarily with the shortest de-

lay. Seizure onset detectors are suited for applications requiring a rapid response to

a seizure, while seizure event detectors are suited for applications requiring an accu-
rate account of seizure activity over a period of time. The following sections present
applications of both detector types and discuss how the specifics of an application
dictate how to balance detection delay against accuracy of detection.

1.2.1 Applications of Seizure Onset Detection

Computerized seizure onset detection can facilitate the initiation of delay-sensitive di-
agnostic, therapeutic, and alerting procedures. Within the realm of diagnosis, seizure
onset detection could be used to quickly initiate functional neuroimaging studies de-
signed to localize the cerebral origin of a seizure. In this case, rapid initiation is
important since the accuracy of such imaging studies diminishes the greater the delay
between seizure onset and infusion of the imaging radiotracer [10]. Within the realm
of therapy, seizure onset detection could be used to trigger neurostimulators designed
to affect the progression of a seizure [60]. In this application, rapid initiation is im-
portant since the likelihood of affecting a seizure seems to decrease the longer the
delay between the onset of a seizure and the start of stimulation [22]. Finally, within
the realm of alerting, seizure onset detection could prompt a patient or care provider
to ensure safety or administer a fast-acting anticonvulsant. In this scenario, waiting
too long to alert the patient or care provider increases the chances that symptoms of
the seizure will leave them unable to respond.

1.2.2 Applications of Seizure Event Detection

Computerized seizure event detectors can enable physicians to better titrate therapy
(pharmacologic or otherwise) over time. Currently, anti-seizure therapy is dispensed
based on a individual's account of the number of and severity of the seizures they
experienced between clinic visits. Unfortunately, too many patients produce an inac-
curate tally [24], which may lead a physician to prescribe too much, too little, or the
incorrect medication. Prescribing too much medication results in toxic side-effects
while prescribing too little means a continuation of frequent seizures. A wearable
device capable of computerized seizure event detection within the ambulatory setting
could provide physicians with a summary of the number, frequency, duration and
time of day an individual experiences seizures. By correlating this information with
different medication regimens a physician could more quickly converge on a treatment
plan that maximally benefits the individual.

1.2.3 Application Dependant Performance

For a seizure onset detector the speed with which a seizure is recognized can be

increased at the expense of seizure detection accuracy. The degree to which one

favors detection speed or accuracy is dictated by the problem to be solved. For

instance, an application such as automatically infusing a radiotracer for the purpose of

a neuroimaging study requires excellent detection accuracy even if it is at the expense

of latency. In contrast, applications such as automatically initiating neurostimulation

call for emphasizing detection speed over accuracy because of the benign effect of
delivering many types of neurostimulation outside of the seizure state [60].

1.3 Why is Seizure Detection Challenging?

Seizure onset and event detection is most often accomplished through analysis of the
Electroencephalogram(EEG). The EEG is a multichannel recording of the electrical
activity generated by collections of neurons within the brain; different channels reflect

the activity within different brain regions. When the EEG is measured using non-
invasive electrodes arrayed on an individual's scalp it is referred to as scalp EEG; and
when it is measured using electrodes placed on the surface of the brain or within its
depths it is referred to as intracranial EEG.
The property of scalp and intracranial EEG that most complicates the seizure
detection task is its variability across individuals with epilepsy [18], both in the
seizure and non-seizure states. Typically, following the onset of a seizure, a set of EEG

channels develops rhythmic activity that reflects underlying neuronal hypersynchrony.

Both the location of the involved EEG channels as well as the spectral content of the

rhythmic activity varies across individuals. Furthermore, the EEG signature of one
patient's seizure may closely resemble the signature of abnormal, non-seizure EEG
gathered from the same patient or from a different patient [46].
Within the scalp EEG the seizure detection task is further complicated by the
physical properties of the signal. The scalp EEG is most sensitive to the activity
of neurons on the brain surface; consequently, the activity of neurons within deep
brain structures has almost no influence on the scalp EEG. When the epileptic neural
network is deep within the brain, the scalp EEG may reflect physical sequelae of
the seizure, such as repetitive eye-blinks (eye flutter) or muscle contractions, before
reflecting hypersynchronous neural activity. Seizures of this type are difficult to detect
with high specificity and low latency since activity such as eye flutter and muscle
contractions are routinely observed as an individual partakes in the activities of daily
life.
Another property of scalp EEG that makes seizure detection challenging is its sus-
ceptibility to contamination by non-physiologic sources. The sway of EEG electrode
cables, alterations in the electrode-skin interface, and the coupling of AC harmonics
from electric machinery can all produce spectral changes that affect the performance
of a seizure detector.

1.4 Related Work

Research into seizure detection methods began with the development of seizure event
detectors [18]. The detectors developed were meant to detect the seizures of any indi-
vidual with epilepsy, i.e., they were patient non-specific. The variability within EEG,
along with the challenges discussed in Section 1.3, severely limited the detection ac-
curacy of these patient non-specific detectors. To improve performance, investigators
developed patient-specific event detectors, i.e., detectors that could be tailored to the
EEG of an individual [44]. These detectors exhibited improved performance because
seizure and non-seizure EEG recorded from an individual exhibits less variability as
shown in Chapter 3. Years later, the development of diagnostic and therapeutic ap-
plications that require initiation following seizure onset motivated the development
of seizure onset detection algorithms.

1.4.1 Scalp EEG Seizure Event Detectors

One of the earliest patient non-specific seizure event detectors was the one developed
by Gotman [18] in 1982. The Gotman algorithm searches for the hallmark sign of
seizures: sustained rhythmic activity. The algorithm sequentially searches a number
of EEG channels for the presence of rhythmic activity with a dominant frequency
between 3-20 Hz and an amplitude at least 3 times greater than that of a background
window; whenever the degree of rhythmicity exceeds a threshold on at least two
channels and persists for 4 seconds a seizure is declared.
The Gotman algorithm successfully detects seizures whose evolution includes sus-
tained rhythmic activity with a fundamental below 20 Hz; it is not successful in
detecting seizures consisting of EEG containing a mixture of frequencies or those
with low amplitude high frequency activity. Since the scalp EEG of individuals with
epilepsy contains pathologic, normal, and artifact-induced bursts of rhythmic activ-
ity, a significant fraction of detections produced by the Gotman algorithm are not
associated with seizures [18]. A recent evaluation of the Gotman algorithm on 652
hours of scalp EEG that included 126 seizures from 28 patients [48] demonstrated
that this approach detects 50% of test seizures and declares 0.5 false detections per
hour.
Since Gotman's work, investigators have developed seizure event detectors that
utilize more sophisticated signal processing to characterize the rhythmicity associated
with seizures as well as more sophisticated schemes for determining whether that
activity is representative of an ongoing seizure. An example of such efforts is the
Reveal seizure detector developed by Wilson [64]. The Reveal algorithm decomposes
2 second EEG epochs from each input channel into time-frequency atoms using the
Matching Pursuit algorithm. Reveal then employs hand-coded and neural network
rules to determine whether features derived from the atoms of a channel are consistent
with a seizure taking place on that channel. The thresholds for some of the neural
network rules are determined using both archetypal seizures from individuals with
epilepsy and background EEG from individuals without epilepsy.
In [64] Wilson reported that the Reveal algorithm detected 76% of 672 seizures
gathered from 426 individuals with epilepsy, and that it declared false detections at
a rate of 0.11 false detections per hour when tested on data from individuals without
epilepsy. When the Reveal algorithm is made patient-specific [65, 66], they reported
that it improved the specificity of the original algorithm from 0.62 false-detection per
hour to 0.34 false detections per hour while improving the sensitivity to 78%. The
Reveal algorithm has a sensitivity superior to that of the classic Gotman algorithm,
but as we show in Chapter 3, it has poor specificity when processing the scalp EEG
of patients with abnormal, non-seizure rhythmic activity.

1.4.2 Scalp EEG Seizure Onset Detectors

Saab developed a patient non-specific seizure onset detector [48]. Saab's algorithm
uses features derived from a wavelet decomposition of each EEG channel to estimate
the probability of a seizure event. Whenever the probability exceeds a user defined
threshold for a given period of time, the algorithm declares the onset of a seizure.
When evaluated on 652 hours of scalp EEG that included 126 seizures from 28 pa-
tients, Saab's algorithm detected 78% of seizures with a median detection latency of
9.8 seconds and a false detection rate of 0.86 false detections per hour. Saab reported
that missed seizures included those with onsets characterized by focal activity, mixed
frequencies, or short duration; and that false detections were mainly caused by short
bursts of rhythmic activity, rapid eye blinking, and chewing.
Qu developed the first patient-specific seizure onset detection algorithm [43, 44,
45, 46]. Qu's patient-specific algorithm relies on a nearest-neighbor classifier to assign
a list of features, or feature vector, to the seizure or non-seizure class. The classifier is
trained on seizure and non-seizure feature vectors from a single individual. The fea-
ture vector consists of measures of the EEG's average amplitude, dominant frequency,
and rhythmicity. The classifier sequentially classifies feature vectors derived from the
available EEG channels, and declares a seizure if the set of positively classified chan-
nels matches half of those chosen by an expert. When tested by Qu on 29.7 hours and
47 seizures from 12 patients, the method detected 100% of seizures with an average

delay of 9.35 seconds and a false alarm rate of 0.03 false detections per hour. The
non-seizure EEG that Qu used to calculate the false detection rate of his algorithm

was formed by concatenating segments of EEG extracted at regular intervals from

several days of data. When compared to Saab's work, Qu's work illustrates that a
patient-specific approach can result in improved sensitivity and specificity, but not
necessarily an improvement in detection latency.
Meier developed a seizure onset detection system that is patient non-specific but

is seizure-specific [35]. Meier grouped seizures in a database into 6 categories based

on the frequency of the dominant rhythm that appears following seizure onset. He
then trained a set of support-vector machines, one for each seizure type, to determine
whether a feature vector extracted from an EEG epoch is consistent with one of the
seizure types. Rather than extract and then classify feature vectors from one channel

after another, Meier extracts a single feature vector that includes the average, across
channels, of signal properties such as the number of zero-crossings, wavelet coefficient

power, and cross-correlation. When evaluated on 91 seizures and 1,360 hours of non-
seizure EEG from 57 patients, Meier's algorithm detected 96% of the test seizures
with an average detection delay of 1.6 seconds and false alarm rate of 0.45 false

detections per hour. Meier's approach depends on the test seizure being a member
of one of the 6 defined categories as well as it being recorded using the same number
and position of channels used to record the training seizures. Seizures whose onsets
lack the development of rhythmic activity and instead reflect physical sequelae of the
seizure, such as eye-flutter, do not fall within the defined categories; consequently,
such seizures will be detected later or not at all.

1.5 Thesis Contributions

In this thesis we develop a patient-specific algorithm for detecting seizure onset that

improves technically upon existing methods in the following ways:

e Enhanced Performance: When trained on 2 or more seizures from the same

 patient and tested on 844 hours of continuous scalp EEG from 23 pediatric
subjects, our algorithm detected 96% of 163 test seizures with a median de-
tection delay of 3 seconds (average 4.6 seconds) and a median false detection
rate of 0.07 false detections per hour (average 0.13 false detection per hour).
When evaluated on the same data set, the Reveal algorithm [64] detected 61%
of seizures with a false detection rate of 1.6 false detections per hour.

Relative to Saab's algorithm [48], our algorithm exhibits a shorter detection

delay, a higher seizure detection rate, and a lower false detection rate.

Relative to Qu [46], our detector exhibits a shorter detection delay, comparable

seizure detection rate, and a higher false detection rate. Unlike Qu's algorithm,
ours does not require an expert to identify relevant EEG channels. Furthermore,
to estimate the seizure detection rate of our method, we used a larger data set
than that used by Qu (163 seizures from 23 patients vs 47 seizures from 12
patients). We also used all 844 hours of non-seizure EEG in order to reliably
estimate the false detection rate of our algorithm. In contrast, Qu tested his
algorithm on 29.7 hours of non-seizure EEG formed by concatenating EEG
segments extracted at regular intervals from several days of data.

Relative to Meier [35], our detector has a longer detection delay, equal seizure
detection rate, and a lower false detection rate. However, unlike Meier's algo-
rithm, ours is not restricted to detecting a defined set of seizure types from a
particular population of individuals with epilepsy.

It is important to note that each of Saab, Qu, and Meier used different data
sets from our own when evaluating the performance of their algorithms.

* Uses Multiple Physiologic Signals: Our method can automatically learn

how to use multiple sources of physiologic information to detect seizures when-
ever the scalp EEG alone is unreliable. This capability is important for the
detection of seizures whose onsets lack the development of rhythmic activity
and instead reflect physical sequelae of the seizure such as eye-flutter, muscle
contractions, or changes in heart rate. Such seizures are not easily grouped into
 the categories defined by Meier [35] and therefore will be detected later or not
at all.

* Minimal User Intervention: Our algorithm does not require a user to define

the values of key algorithm parameters. In our approach, a user only needs

to define the onset of activity associated with a seizure in a set of physiologic

signals. The relationships between these signals that distinguish the seizure and
non-seizure periods are automatically learned.

* Suitable for Implantable Medical Devices: Our algorithm can be adapted

for the detection of seizure onset within intracranial EEG, and can be imple-
mented on the low-power hardware of an implantable neurostimulator. When
evaluated on 81 hours of intracranial EEG containing 61 seizures and gath-
ered from 17 adult subjects, our algorithm detected 60/61 seizures within 9.3

seconds, declared a total of 28 false detections, and consumed 12pA when im-
plemented on the hardware in [2]. When evaluated on the same data set and
hardware, a patient non-specific algorithm based on [38] detected 41/61 seizures
within 18.7 seconds, declared a total of 17 false detections, and consumed 32pA.

In this thesis we also discuss the application of our patient-specific algorithm to

the following delay-sensitive therapeutic and diagnostic applications:

* Non-invasive Closed-Loop Control of the Vagus Nerve Stimulator:

Using our seizure onset detector we designed and clinically evaluated the first
non-invasive system that initiates Vagus Nerve Stimulation (VNS) in response
to detecting the onset of a seizure using multiple physiologic signals [57]. As an
example of the system's capabilities, during an 81 hour clinical test of the system

on a patient, the computerized system detected 5/5 seizures and initiated VNS
within 5 seconds of the appearance of ictal discharges in the EEG.

* Computerized Initiation of Ictal SPECT Studies: Using our seizure onset

detector we designed and clinically evaluated a system for initiating a functional

neuroimaging study following seizure onset. The neuroimaging modality, ictal

 SPECT, is used to radiographically localize the cerebral origin of a seizure. Our
system could initiate injection of the radiotracer used for ictal SPECT within
19.3 ± 2.3 seconds in 8/8 prospective trials, while the clinical team required
27.7 + 8.5 seconds, and failed to initiate ictal SPECT in one of the trials.

1.6 Thesis Outline

This thesis is organized as follows: In Chapter 2 we provide background material on
seizures and both the scalp and intracranial electroencephalograms. Next, in Chapters
3-5, we develop and analyze the performance of a machine-learning based, patient-
specific algorithm for the detection of seizure onsets within scalp EEG. In Chapter
6, we illustrate how the detector in Chapter 3 can be extended with information
from other physiological signals in order to detect seizures whose onset does not
immediately involve the development of rhythmic activity within the scalp EEG.
We then present applications of our seizure detection methodology. In Chapter 7
we illustrate how the methods developed in Chapters 3 and 6 were integrated into a
real-time system that initiates vagus nerve stimulation in response to detecting the
onset of a seizure. Chapter 8 illustrates how the method developed in Chapter 3 was
used in a real-time system that infuses the radiotracer used in ictal SPECT following
seizure onset detection. Finally, in Chapter 9, we adapt the feature extraction and
classification stages of our detector so that it may be embedded within a low-power,
implantable medical device.
Chapter 2

Seizures and the

Electroencephalogram

This chapter reviews the pathophysiology underlying seizures as well as their clin-
ical manifestation and categorization. This chapter also reviews properties of the
electroencephalogram and the electrographic characteristics of seizures.

2.1 Epileptic Seizures

Neurons are cells within the brain capable of generating, propagating, and process-
ing electric signals. Neurons connect to other neurons in order to form functional
networks, and the brain can be viewed as a collection of interacting neural networks.
The inputs to a neural network can be excitatory or inhibitory. Excitatory inputs
promote activity among neurons within a network and inhibitory inputs suppress it
[30].
Epileptic seizures are transient periods involving the hyperactivity and hypersyn-
chronization of a large number of neurons within one or more neural networks. These
transient states arise because of a perturbation that creates an imbalance favoring the
excitation of a neural network over its inhibition. The imbalance may arise because
of defects within a neuron, such as an ion channel dysfunction; defects in connections
between neurons, such as deficient inhibitory neurotransmitter synthesis; or defects in
neural network organization, such as the formation of aberrant excitatory connections
between neurons. Defects within neurons, neuronal connections, or neural network
orginization may result from a genetic disorder or from trauma to the central nervous
system during life.
Epileptic seizures are broadly classified according to their cerebral site of origin
and spread. Focal seizures arise from a localized region of the brain's cortex and have
clinical manifestations that reflect that region of the brain. As an example, a focal
seizure originating in the temporal lobe, the part of the brain that processes emotions
and short-term memory, may result in feelings such as euphoria, fear, and deja vu or
hallucinations of taste or smell. Focal seizures may spread to involve other regions of
the brain or the entire brain. As an example, a seizure originating in the left motor
cortex may result in jerking movements of the right upper extremity. When the
seizure spreads to adjacent areas and then the entire brain, whole-body convulsions
ensue.
Generalized seizures begin with abnormal electrical activity that appears to en-
compass the entire cerebral cortex. The manifestations of such widespread abnormal
electrical activity often includes the loss of consciousness. Motor manifestation of
these seizures may include whole-body rigidity and jerking (generalized tonic-clonic
seizure) or whole-body loss of muscle tone (atonic seizure). A seizure that begins
focally and then generalizes is referred to as a secondarily generalized seizure.

2.2 Scalp Electroencephalogram

The scalp electroencephalogram (scalp EEG) is a non-invasive measure of the elec-

trical potentials generated by the activity of tens of millions of neurons within the
brain. The scalp EEG is usually measured through electrodes that are symmetrically
arrayed on the scalp as shown in Figure 2-1. An EEG signal, or channel, is formed
by taking the difference between potentials measured at two electrodes. For exam-
ple, the channel FP1 - F7 is formed by taking the difference between the potentials
measured at the electrodes FP1 and F7. Each EEG channel summarizes activity
localized within a region of the brain; for instance, the channel FP1 - F7 reflects
neural activity originating within the frontal lobe of the left hemisphere. The onset
of a focal seizure involves a change in activity on the few scalp EEG channels that
lie above or near the site of the brain giving rise to a seizure; on the other hand, the
onset of a generalized seizure involves activity on all scalp EEG channels.
The physics of EEG generation constrains both the origin and characteristics of
neural activity visible within the scalp EEG. In particular, the neurons that con-
tribute the most to the scalp EEG are those closest to the scalp surface; in contrast,
the activity of neurons buried within deep brain structures is not observable. Further-
more, the cerebrospinal fluid and skull surrounding the brain act as attenuators that
greatly diminish the amplitude of higher frequency neural oscillations. An important
consequence of these physical limitations is that certain types of seizures, namely
those involving a small, deep region within the brain cannot be observed using the
scalp EEG.

FP

P7 P3 Pz P4 PS

01 02

Figure 2-1: EEG electrodes arrayed symmetrically across the scalp provide a temporal
and spatial summary of the synchronous firing of tens of millions of neurons within
the brain.

Electroencephalographers describe EEG activity in terms of its spatial distribution

on the scalp (frontal, posterior, lateral, and bilateral) as well as its dominant frequency
component. An EEG wave is classified as having a delta component if its dominant
frequency component f is < 4 Hz, a theta component if 4 < f < 8 Hz, an alpha
component when 8 < f < 12 Hz, a beta component when 12 < f < 30 Hz, or a
gamma component when f _ 30 Hz. For example, the alpha wave in EEG parlance
refers to a 10 Hz rhythm that appears most prominently on posterior channels when
a subject closes their eyes and relaxes.
Scalp EEG activity is modulated by the state of vigilance of an individual. In
particular, the dominant frequency and spatial distribution of EEG activity during the
awake state is different than that during sleep. As an example, Figure 2-2 illustrates
awake EEG activity interrupted by an eye-blink at 37 seconds. The awake EEG
background is primarily composed of low frequency activity. The eye-blink results in
a downward deflection of the EEG signal on the channels {FP1-F7,FP1-F3,FP2-
F4, FP2 - F8}. Figure 2-3 illustrates EEG activity recorded during sleep. The 11
Hz oscillation, observed most prominently on the channel FP2 - F4 between 12-14
seconds, is known as a sleep spindle.

T7-P7

P3-01
FP2-F4 I M " 'A I 1

C4-P4
P -02
FP2-FS 4.. . 1P.-
T8-Pa I 1 - "
P8-02

FZ-CZ
CZ-PZ
33 34 35 36 37 38 39 40 41 42 43

Figure 2-2: Example of 10 seconds of awake EEG interrupted by an eye-blink at

37 seconds. The eye-blink results in a downward deflection on the EEG channels
{FP1 - F7,FP1 - F3, FP2- F4, FP2 - F8}.

Scalp EEG is also easily corrupted by both physiological and non-physiological

artifacts. Physiological artifacts include sweat, chewing, eye-blinks, and scalp mus-
cle contractions. Non-physiological artifacts include power-line noise and EEG cable
motion. As an example, Figure 2-4 illustrates how chewing affects the scalp EEG.
Chewing results in rhythmic, high-frequency activity that is most prominently ob-
served on channels on either side of the scalp (e.g. F7 - T7 and F8 - T8).
 FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3 ," III F I
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4 ' I
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ

Figure 2-3: Example of 10 seconds of sleep EEG interrupted by 11 Hz oscillations

known as sleep spindles. The sleep spindles are most visible on the channel FP2- F4
between 8-10 and 12-14 seconds.

FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
533 534 535 536 537 538 539 540 541 42 43

Figure 2-4: Example of 10 seconds of awake EEG interrupted by the rhythmic, high-
frequency activity associated with chewing. The high-frequency activity caused by
chewing can be seen on the EEG channels F7 - T7 and F8 - T8.
2.2.1 Seizures Within the Scalp Electroencephalogram

Within the scalp EEG, seizures manifest as a sudden redistribution of spectral energy
on a set of EEG channels. The spectral energy redistribution is caused by hypersyn-
chrony of neurons within an epileptic neural network, and consists of an appearance

or disappearance of frequency components within the 0-25 Hz band [19]. However,

which spectral components vanish or rise to prominence varies across patients. Fur-
thermore, the EEG channels demonstrating the spectral energy change also varies
across patients since it is a function of the cerebral site of origin of a seizure.

As an example, Figure 2-5 and Figure 2-6 illustrate seizures from patients A and
B respectively. The seizure in Figure 2-5 begins at 1723 seconds and consists of
flattening of the EEG signal across all channels followed by the appearance of a beta

band rhythm on the channels {F3 - C3, C3 - P3}. Then, over the course of a few
seconds, the amplitude of this rhythm increases as its frequency decreases and settles
within the theta band.

FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
19 1721 3 1725 1727 1729 1731 1733 1735

Figure 2-5: Example of a seizure within the scalp EEG of Patient A. The seizure,
which begins at 1723 seconds, involves flattening of the EEG signal across all channels
followed by the appearance of a beta band rhythm on the channels {F3-C3, C3-P3}.

The seizure in Figure 2-6 begins at 6313 seconds with the onset of a theta band
rhythm that is most prominent on the channels {F7 - T7, T7 - P7}. Other EEG
channels also exhibit a change following seizure onset. The channel {C3 - P3} de-
velops a theta band rhythm while the channel {FP2 - F8} develops a delta band

rhythm.

FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
6307 6309 6311 6313 6315 6317 6319 6321 6323

Figure 2-6: Example of a seizure within the scalp EEG of Patient B. The seizure,
which begins at 6313 seconds, involves the appearance of a theta band rhythm on the
channels {F7 - T7, T7 - P7}.

Between seizures, the EEG of an individual with epilepsy may exhibit abnormal
rhythmic activity or discharges. The spatial and spectral characteristics of these dis-
charges varies across patients. Figure 2-7 illustrates an abnormal discharge within the
EEG of Patient A. The discharge, which involves most EEG channels, falls between
2884-2892 seconds, and is characterized by a repeating pattern of high-amplitude
spikes followed by broad waves. While these discharges are seen frequently in the
awake EEG of Patient A, they are not accompanied by the physical symptoms as-
sociated with Patient A's seizure (Figure 2-5). Consequently, a detector designed to
react to the seizures of Patient A should not produce an alarm upon the onset of one

of these discharges. In another patient, this type of activity may be associated with

physical symptoms, as is visible in the seizure illustrated in Figure 5-25.

Figure 2-8 illustrates abnormal rhythmic activity observed within the EEG of
 FP1-F7
F7-T7
T7-P7
P7-01
FPI-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
2878288 2 84 28 2888 289 2892

Figure 2-7: Example of an abnormal discharge within the scalp EEG of Patient A.
The discharge, which occurs between 2884-2892 seconds, involves a repeating pattern
of high-amplitude spikes followed by broad waves that can be seen on most EEG
channels.

Patient B. In this case, a theta band rhythm is observed on the frontal channels
{FP1 - F3, F3 - C3} during the awake state between 6126-6130 seconds. This

activity is both frequent and abnormal, but it is not accompanied by the physical
symptoms associated with Patient B's seizure (Figure 2-6). Once again, a detector
designed to react to the seizures of Patient B should not produce an alarm upon the
onset of this rhythmic activity.

The previous examples illustrate the variability in the spectral and spatial signa-
ture of seizure and non-seizure activity across patients. This variability is the primary
reason why patient non-specific detectors exhibit poor sensitivity and specificity for
seizure events [18]. In contrast, for a given individual, seizures that emerge from the
same cerebral site exhibit similar clinical symptomatology and electrographic spatial
and spectral characteristics. As an example, Figures 2-9 and 2-10 illustrate more

seizures recorded from patients A and B respectively. Note the similarity in the spa-
tial and spectral character of these seizures and those illustrated in Figures 2-5 and
2-6.

The long-term stability of an individual's seizure signature has not been studied
 FP1-F7
F7-T7
T7-P7 y~3~
J*~s~j"?Lrr~c~/)~
Y
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4 WI
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
6126 6128 6130 6132 6134 6136 6138 6140 6142

Figure 2-8: Example of abnormal rhythmic activity within the scalp EEG of Patient
B. The rhythmic activity, which occurs between 6126-6130 seconds, involves a theta
band rhythm on the frontal channels {FP1 - F3, F3 - C3}.

FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ )SI

Figure 2-9: A second seizure within the scalp EEG of Patient A. The second seizure,
which begins at 6210 seconds, resembles the seizure shown in Figure 2-5.
 FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
2375 2377 2379 2381 2383 2385 2387 2389 2 91

Figure 2-10: A second seizure within the scalp EEG of Patient B. The second seizure,
which begins at 2381 seconds, resembles the seizure shown in Figure 2-6.

rigorously. However, in our experience, seizures recorded months apart from patients

without a progressive brain disorder exhibit very similar spectral and spatial char-
acteristics. As an example, Figures 2-11 and 2-12 illustrate two seizures that were

recorded approximately 10 months apart from Patient C. Both seizures begin with
rapid eye-blinking, at 1488 seconds and 992 seconds respectively, which manifests in
the EEG as high-amplitude, downward deflections on the channels {FP1- F3, FP2-
F4, FP1 - F7, FP2 - F8}. Later, both seizures exhibit a 3-4 Hz theta wave most
prominently on the EEG channel T4 - T6 at 1492 seconds and 998 seconds respec-
tively.

2.3 Intracranial Electroencephalogram

Similar to scalp EEG, the intracranial Electroencephalogram (iEEG) provides a spa-

tial and temporal summary of the electrical activity of a population of neurons. Since
intracranial EEG electrodes are placed on the brain cortex or deep within brain struc-
 F3-C3
C3-P3
P3-01

Fp2-F8
F4-C4 I I

C4-P4 H t wv o t - T6.FAA
P4-02
Fp-F7
F7-T3
C3-T5

Fp2-F8
F8-T4
T4-T6
T6-02

1484 1486 1488 1490 1492 1494 1496 1498 1500 1502 1504 1506 1508 1510

Figure 2-11: Example of a seizure within the scalp EEG of Patient C. The seizure,
which begins at 1486 seconds, involves rapid eye-blinking followed by the appearance
of a 3-4 Hz theta wave on the channel T4 - T6.

Fpl-F3
F3-C3
C3-P3

P3T5-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5
T5-O1
Fp2-F8
F8-T4
T4-T6
T6-02

988 990 992 994 996 998 1000 1002 1004 1006 1008 1010 1012 1014

Figure 2-12: Example of a seizure recorded from Patient C approximately 10 months

after recording the seizure in Figure 2-11.
tures, the intracranial EEG signal summarizes the activity of a smaller population of
neurons and is immune to corruption by many of the physiologic and environmental
artifacts that plague scalp EEG. Because they are highly invasive, iEEG electrodes
can only be implanted in a limited number of brain sites at any given time; so in-
tracranial EEG offers higher spatial resolution but spatial coverage that is worse than
that of scalp EEG.
The higher spatial resolution of iEEG allows one to notice the neuronal hyper-
synchrony associated with a seizure tens of seconds before the same phenomenon is
noticeable within the scalp EEG [39]. At the same time, the higher spatial resolution
of iEEG permits the recording of a wider gamut of abnormal, non-seizure activity
that is not visible within the scalp EEG [26, 59, 62]. This activity can confound the
seizure detection process.

2.3.1 Seizures Within the Intracranial Electroencephalogram

The manifestation of seizures within the intracranial EEG involves a sudden redis-
tribution of spectral energy on a set of iEEG channels. The spectral energy change
typically consist of an appearance or disappearance of frequency components within
0-65 Hz band [20]. This frequency range is wider than that associated with scalp
EEG because of the absence of the attenuating effect of the cerebrospinal fluid and
skull. Prior to the spectral energy change, discrete events such as a spike or train of
spikes may appear.
As was the case with scalp EEG, the manifestation of seizures within the iEEG
varies significantly across patients. To observe this variability consider Figure 2-
13. The top panel corresponds to a seizure from Patient D while the bottom panel
corresponds to a seizure from Patient E. The seizure of Patient D begins at 5 seconds
and consists of a few spikes that transition into a high-amplitude spike train with a
period of 1 Hz. The seizure of Patient E also begins at 5 seconds and consists of a
few spikes followed by low-amplitude, high-frequency activity.
Outside the context of a seizure the intracranial EEG of an individual with epilepsy
may contain a wide variety of rhythmic activity with characteristics that resemble
 1000

800

400
200-
0
-200

0 5 10 15 20 25 30 35 40 45
Time (Seconds)

800

600

400

200

-200

0 5 10 15 20 25 30 35
Time (Seconds)

Figure 2-13: Example of seizures within the iEEG of patients D (top panel) and C
(bottom panel). The seizure of Patient D begins at 5 seconds and consists of a few
spikes that evolve into a high-amplitude spike train. The seizure of Patient E begins
at 5 seconds and consists of a few spikes followed by low-amplitude, high-frequency
activity.

those of a seizure. As an example, consider Figure 2-14. The top panel of Figure
2-14 illustrates a seizure recorded from Patient F while the bottom panel illustrates
a burst of rhythmic activity recorded from the same patient. An algorithm sensitive

only to changes in spectral energy without regard to the frequency range or channels
on which those changes occur [38] may falsely declare the activity in the bottom panel
as a seizure.

For a given patient, the spectral character of seizures within the iEEG is similar
from one seizure to the next provided that the seizures originate from the same
cerebral site [14, 29]. As an example, consider the two seizures from Patient D
illustrated in Figure 2-15. Both seizures begin with a few spikes that transition into
a high-amplitude spike train with a period of 1 Hz. However, the spatial character
of seizures within the iEEG, that is which iEEG channels are involved, tends to

vary more than in the setting of scalp EEG. More specifically, for a given patient,

neighboring iEEG channels may alternate in exhibiting the first signs of iEEG activity

associated with the onset of a seizure from the same brain region.
 Time (Seconds)

-""" L
4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025
Time(Seconds)

Figure 2-14: The iEEG of Patient F contains rhythmic activity that is associated
with a seizure (top panel) as well as rhythmic activity that is not associated with any
clinical symptoms (bottom panel).

IUUr

1000

800
600
400
200
0

-200

0 5 10 15 20 25 30 35 40 45
Time (Seconds)

15 20
Time (Seconds)

Figure 2-15: Example of two seizures within the iEEG of Patient D. Both seizures
begin at 5 seconds and consist of a few spikes that evolve into a high-amplitude spike
train.
2.4 Summary

Seizures are transient periods of neural network hyperactivity and hypersynchrony.

The clinical symptoms that accompany a seizure depend both on its cerebral site of
origin as well as its pattern of spread to surrounding brain regions. The scalp and
intracranial electroencephalograms, which measure the aggregate electrical activity
of populations of neurons, can be used to detect seizures. The scalp EEG offers poor
spatial resolution but high spatial coverage while the intracranial EEG offers high
spatial resolution but less spatial coverage.
Within both the scalp and intracranial electroencephalograms seizures manifest
as a redistribution of spectral energy on a set of channels. The manner with which
spectral energy is redistributed varies significantly across patients. For a given indi-
vidual, seizures from the same cerebral site generally exhibit stereotyped onset and
evolution.
54
Chapter 3

Patient-Specific Seizure Onset

Detection using the Scalp EEG

In this chapter we present the architecture of a novel algorithm for detecting the onset
of a seizure within the scalp EEG. The algorithm treats the seizure onset detection
problem as a binary classification task that is to be solved within a patient-specific
context. The feature vector used by the algorithm encodes the time evolution of
spectral and spatial properties of the scalp EEG.

3.1 Overview of Binary Classification

Binary classification is the process of assigning an observation to one of two classes or

categories in a manner that optimizes a chosen performance objective; for instance,
minimizing the probability of an erroneous classification. As an example of binary
classification, consider the problem of determining the absence or presence of an
aircraft using returned radar pulses. In this scenario the observation is the radar
pulse, which represents the presence (Class C1 ) or absence (Class C2) of an aircraft.
In the seizure onset detection problem the observation is an EEG epoch and the two
classes are seizure activity (Class Ci) and non-seizure activity (Class C2).
Determining the class membership of an observation involves two steps. First,
salient properties or features that most discriminate between instances of each class
are extracted from the observation and assembled into a vector of features. In the
case of radar signal processing, the features may be the energy within a series of
returned radar pulses. Next, a classifier that is trained to recognize the difference
between feature vectors extracted from instances of each class determines the class
membership of the observation based on its associated feature vector.
The success of a binary classification task depends strongly on which features are
extracted from an observation as well as on the classifier used to determine class
membership. Features that have a distribution of values for observations belonging
to class C1 and another, very different distribution for observations belonging to class
C2 result in good performance. A classifier capable of deducing from training data
a simple rule that accurately distinguishes between feature vectors extracted from
instances of each class is also important to good performance.

3.2 Feature Vector Design

Seizures are dynamic processes. The onset of a seizure induces within the scalp EEG
a sudden change in the spectral energy distribution of a set of EEG channels. As the
seizure progresses, the spectral and spatial character of the EEG continues to evolve.
The following sections detail how the feature vector used in our detector is designed
to encode the time evolution of spectral and spatial properties within the scalp EEG.

3.2.1 Spectral Features

Following the onset of most seizures, scalp EEG channels that record brain activity
within regions involved in the seizure exhibit rhythmic activity. The spectral structure
of this rhythmic activity may be composed of multiple frequency components. As an
example, Figure 3-1 illustrates a seizure that begins following 2994 seconds, and
Figure 3-2 illustrates the frequency spectrum of the channel FP2- F4 between 2994-
2999 seconds. For this seizure, the rhythmic activity on the channel FP2 - F4 is
composed primarily of frequency components at 2, 5, and 11 Hz.
Considering several spectral components that constitute an EEG epoch, and not
 FP1-F7
F7-T7 I VIII A \
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
FS-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ

Figure 3-1: Example of a seizure whose onset is associated with rhythmic activity
that contains a mixture of prominent frequency components. Seizure onset follows
2994 seconds and mostly involves channels on the right side of the head (channels
with even numerals).

i6 100

0 50

10 12 14
Frequency (Hz)

Figure 3-2: Frequency spectrum of the channel FP2 - F4 following the onset of the
seizure illustrated in Figure 3-1. The spectrum contains large spectral components
at 2, 5, and 11 Hz.
simply the dominant spectral component as was done in early seizure detection al-

gorithms [18], is important in detecting seizure onsets with high specificity. The

dominant frequency of a seizure onset epoch may overlap the dominant frequency of

an epoch of non-seizure activity, but what differentiates the two is the presence or
absence of less dominant spectral components. The following examples illustrate this
point.

Figure 3-3 illustrates the spectrum of an eye-blink superimposed on the seizure

spectrum (Figure 3-2). The two spectra overlap in the frequency range 0-4 Hz, but

the spectrum of the seizure activity contains an 11 Hz component that is absent from
the spectrum of the eye-blink.

200
- v Seizure Onset
tEye Blink
150

S100

50

0-

-50
0 2 4 6 8 10 12 14 16 18 20 22 24
Frequency (Hz)

Figure 3-3: Superposition of the frequency spectra of an eye-blink and the rhythmic
activity observed on the channel FP2 - F4 following the onset of the seizure in
Figure 3-1. The seizure spectrum contains an 11 Hz component that is absent from
the spectrum of the eye-blink

Figure 3-4 illustrates the spectrum of a sleep spindle superimposed on the seizure

spectrum. In this case, the two spectra overlap in the frequency range 10-12 Hz, but
the seizure spectrum contains larger low-frequency spectral components.

Finally, Figure 3-5 illustrates the spectrum of chewing superimposed on the seizure

spectrum. The two spectra overlap within the frequency range 0-12 Hz, but the seizure

spectrum contains less high-frequency content relative to the chewing spectrum.

 100

50
0 2 4 6 8 10 12 14 16 18 20 22 24
Frequency (Hz)

Figure 3-4: Superposition of the frequency spectra of a sleep spindle and the rhythmic
activity observed on the channel FP2- F4 following the onset of the seizure in Figure
3-1. The seizure spectrum contains low-frequency spectral components that are larger
than those in the spindle spectrum.

200
a- Seizure Onse
cChewing

150

100

S50

0-

-50
0 2 4 6 8 10 12 14 16 18 20 22 24
Frequency (Hz)

Figure 3-5: Superposition of the frequency spectra of chewing and the rhythmic
activity observed on the channel FP2 - F4 following the onset of the seizure in
Figure 3-1. The seizure spectrum contains less high-frequency content relative to the
chewing spectrum.
 Our feature vector encodes the spectral structure of an L second epoch from EEG
channel k by passing the epoch through the M-band filterbank shown in Figure 3-6,
and then measuring the energy in the subband signals Si, i = 1,..., M. The energy

in the subband signal Si from channel k is represented by the feature Ci,k. The
filterbank spans the frequency range 0.5-24 Hz since within this range one observes
most physiologic and pathophysiologic scalp EEG activity [19].

EEG Channel k ...

4-

S1 S S3 SM

Compute Energy in Subband Signal Si

X1,k X2,k X3,k XM,k

Figure 3-6: M-band filterbank that measures the energy within the spectral compo-
nents of an L second epoch taken from a single EEG channel.

As an example, Figure 3-7 illustrates how energies within the frequency bands
defined by an M = 8 filter filterbank can differentiate between the spectra of seizure
and non-seizure activity. The energy within the frequency bands defined by the
rightmost three filters (thick line) differentiate between the seizure spectrum (red)
and the spectrum associated with chewing (blue).

3.2.2 Spatial Features

The set of EEG channels on which EEG activity is observed, which is referred to as
its spatial distribution, can be important in differentiating between seizure onset and
non-seizure activity.
 180
I % I I . '

SI I

120 i-

n
60 I
in Iml
I
u ,- •
' • • i "n v '
m\
I ii

0 2 4 6 8 10 12 14 16 18 20
Frequency (Hz)

Figure 3-7: Energy within the frequency bands defined by the rightmost three filters
(thick line) of the filterbank differentiates between the seizure spectrum (red) and
chewing spectrum (blue)

FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
2989 2990 2991 2992 2993 2994 2995 29 6 2997

Figure 3-8: The spatial distribution of EEG activity can be used to differentiate
seizure from the non-seizure activity. The sleep spindles between 2989-2992 seconds
do no involve activity on the channels C4 - P4 and T8 - P8. In contrast, the seizure
activity following 2994 seconds involves both of these channels.
 As an example, Figure 3-8 illustrates sleep spindles between 2989-2992 seconds,

and seizure activity following 2994 seconds. The sleep spindle and seizure activity are
both rhythmic, but the seizure involves EEG channels (C4 - P4 and T8 - P8) that
are not involved in the sleep spindle event. Consequently, the absence of activity on
a set of channels during a non-seizure event and its presence during seizure onset can
serve to differentiate EEG activity beyond what is possible using spectral information
alone.

To simultaneously capture the spectral and spatial information contained within

an L second EEG epoch at time t = T, we concatenate the M spectral energies

extracted from each of N EEG channels to form a composite feature vector XT with
M x N elements. The process of generating XT is illustrated in Figure 3-9. It is
important to note that the feature vector XT automatically captures relationships

between the spectral structure of different EEG channels. Previous algorithms relied
on experts to define those relationships [44, 46].

Spectral Features Spectral and Spatial

Features

~x1,1

Xl,N
EEG Ch N ... . x

- .. . X2,N

Figure 3-9: Formation of an intermediate feature vector XT that captures the spectral
and spatial properties of an epoch at time t = T
3.2.3 Time Evolution

The feature vector XT developed in section 3.2.2 uses all EEG channels to capture

the spatial and spectral content of an EEG epoch at time t = T, but it does not
capture how the current epoch relates to those in the recent past. Consequently, the

feature vector XT cannot represent how a seizure emerges from the background EEG
nor how it evolves during the onset period. In order to capture the relation between
feature vectors across time we form a composite feature vector XT that is the result

of concatenating feature vectors from W non-overlapping epochs. This is similar to

what was done in [7] for off-line intracranial seizure onset detection. The feature

vector XT is constructed as shown in equation 3.1, where XT is a spectral and spatial

feature vector extracted from an epoch of length L seconds at time t = T.

XT = [X,_T(wL ... XT-2L XT-L XT] (3.1)

Encoding the temporal evolution of EEG through the formation of Xr is not

analogous to forming a single feature vector XT using a longer epoch length L. In the

former approach, discrete events are preserved and in the latter the spectral signature
of discrete events within the longer epoch is smeared.

To appreciate the value of capturing the temporal evolution of the feature vectors

XT consider deriving the composite feature vector XT for the seizure shown in Figure

3-10 using only the channel F3 - C3 (recall that in the actual detector all channels
are used). In this example the epoch length used to from XT is L = 2 seconds long
and the feature vector XT is formed by concatenating W = 3 feature vectors: XT-4,

XT-2, and XT.

The seizure in Figure 3-10 involves a sequence of events that can be modeled
through the feature vector XT. The seizure begins at 1723 seconds with a spike

followed by a period of low amplitude EEG known as an electrodecrement. Next, a

beta band rhythm appears most prominently on the channel F3 - C3, and, over the

period of a few seconds, increases in amplitude and decreases in frequency towards

 ~ -~I V~-[-
V r(~n;r~~rVlr *F~(VrPr\CYliNYlr(l~W*n\rYIWrrV(
vI
1~Vh~CVJ~"Cy~LUy~,k~
i I I
"C~VJ~H\~L4hChthCsln*~4p)rS1YY~S~
cL~r/rc~
ch
~vLy
~-~JhrULYCJ. I -IL-l i II-n~-- L ~G i

1733 1734

Figure 3-10: Scalp EEG seizure involving a sequence of discrete spectral events. The
first event, at 1723 seconds, is a spike involving all EEG channels. Next, all EEG
channels exhibit a period of low-amplitude EEG. Finally, at 1725 seconds, a beta
band rhythm that increases in amplitude and decreases in frequency appears on the
channel F3 - C3.

the theta band.

To more clearly demonstrate the series of spectral events that constitute the seizure

in Figure 3-10, we show the spectrogram of channel F3 - C3 in Figure 3-11. A

spectrogram is a plot of the spectral content of a signal as function of time. In a
spectrogram spectral components with greater energy are shown in darker colors and

those with less energy in brighter colors. In Figure 3-11 one can see a high energy

event with spectral content up to 30 Hz between 3.5-4.5 seconds, this corresponds

to the spike at 1723 seconds in Figure 3-10. Between 4.5-6.5 seconds a decrease in
energy (brighter colors) in the frequency band 0-10 Hz is observed, this corresponds
to the electrodecrement between 1723-1725 seconds in Figure 3-10. Finally, between
6.5-13.5 seconds one observes a negatively sloped, dark colored line representing a
spectral component that begins with a frequency of 20 Hz and decreases towards a

frequency of 5 Hz. This corresponds to the onset and evolution of the beta band

rhythm at 1725 seconds in Figure 3-10.

Figures 3-12 and 3-13 illustrate how the composite feature vector XT is able to
model the sequence of events associated with both seizure onset and evolution. In

Figure 3-12 note how XT, which is composed of the triplet of feature vectors XT- ,
4
 S20
u_
15

10
5

8. 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5 9.5 10.5 11.5 12.5 13.5
Time (Seconds)

Figure 3-11: Spectrogram of Channel F3 - C3 illustrating sequence of events that

compose the onset and evolution of the seizure in Figure 3-10.

ZT

56
50
46
40
36
6-30
26
Cr
e20
U.

1.5 2.5 3.5 4.5 6.5 6.5 7.5 8.5 9.5 10.5 11.5 12.5 13.5
Time (Seconds)

Figure 3-12: Spectrogram of channel F3 - C3 illustrating how the feature vector

XT models the transition from background to seizure onset. The feature vector XT
captures the period of low amplitude activity, XT-2 captures the spike, and XT-4
captures the background EEG activity preceding the spike.
 Zr

66

15
530

10

85 1.5 2 3,5 45 565 65. 7.5 8.6 9.5 10.5 11.5 12.5 13.6
Time (Seconds)

Figure 3-13: Spectrogram of channel F3 - C3 illustrating how the feature vector XT

models the evolution of a seizure. The feature vector XT captures the beta band
rhythm, XT-2 captures the period of low amplitude activity, and XT-4 captures the
spike.

XT-2, and XT spans the transition from background to seizure activity. The feature
vector XT captures the period of low amplitude activity, XT-2 captures the spike,
and XT-4 captures the background EEG activity preceding the spike. When shifted
forward in time as shown in Figure 3-13, the composite feature vector X7 models
the seizure's evolution. The feature vector XT captures the beta band rhythm, XT-2

captures the period of low amplitude activity, and XT-4 captures the spike.

3.3 Feature Vector Classification

To classify the feature vector XT as representative of seizure or non-seizure activity

we used the Support-Vector Machine (SVM) algorithm [9]. The SVM algorithm uses

training seizure and non-seizure feature vectors to learn a decision boundary that

separates these two classes of activity. Once the decision boundary is learned, the

SVM algorithm determines the class membership of a newly observed feature vector

Xr based on which side of the boundary the vector falls.

In its simplest form, the SVM algorithm determines a decision boundary in the
form of a high-dimensional plane, or hyperplane, in the space of features. The hy-
perplane is chosen to maximize the classification margin, which is the geometric
distance between the hyperplane and the boundary cases of each class [61]. The
boundary cases are known as support-vectors. If the classes cannot be well separated
by a hyperplane, as is the case in the seizure onset detection problem, the SVM can
be used to determine a quadratic or circumferential decision boundary. Conceptually,

the SVM determines a nonlinear boundary in the input feature space by solving for
a linear boundary in a higher-dimensional feature space induced by a kernel.

0.8 0% o o o
000
N
S0
50 OO xx
0r 4
0.63

I. 1

0.2

0 0.2 0.4 0.6 0.8 1

Energy (0-16Hz)

Figure 3-14: Two-dimensional seizure (red) and non-seizure (blue) feature vectors
extracted from a single iEEG channel. The first dimension of this feature space
is defined by the energy in the spectral band 0-16 Hz and the second dimension
corresponds to the energy in the 25 + 11 Hz band. The seizure vectors extracted from
the first 7 seconds of the seizure are numbered 1-5.

As an example, Figure 3-14 plots, in a two-dimensional feature space, iEEG seizure

feature vectors (red circles) and non-seizure feature vectors (blue crosses) extracted
from a single iEEG channel. The first dimension of this feature space is defined by
the energy in the spectral band 0-16 Hz and the second is the energy in the 25 ± 11
Hz band. The seizure feature vectors (red circles) form two clusters: one cluster with

fewer feature vectors that lies in close proximity to the non-seizure feature vectors,
and a second cluster with a much greater number of feature vectors that lies far away
from the non-seizure vectors. The seizure vectors closest to the non-seizure vectors
are associated with the seizure onset period and those further away are associated
with later stages of the seizure. In Figure 3-14, the seizure vectors numbered 1-5 are
extracted from the first 7 seconds of a seizure.
Figure 3-15 shows the linear decision boundary separating the iEEG seizure feature
vectors from the non-seizure vectors. As can be seen the two classes are not well
separated by a linear decision boundary as there are numerous crosses on the side of
the hyperplane designated for seizure feature vectors.

0.8 0% o o o
x
OO OO x x
5 00
O 4

0.2 NN
2
N *

0 0.2 0.4 0.6 0.8 1

Energy (0-16Hz)

Figure 3-15: Linear decision boundary separating seizure (red) and non-seizure (blue)
feature vectors extracted from a single iEEG channel. The decision boundary was
determined using the SVM learning algorithm.

Figure 3-16 shows a nonlinear decision boundary separating the same iEEG seizure
and non-seizure feature vectors. The nonlinear boundary, which is computed using
a radial basis kernel, is able to better separate the seizure and non-seizure feature
vectors.
The earliest seizure vector that is correctly classified by either the linear or non-
linear decision boundary is the third seizure vector. The SVM learning algorithm
allows the user to specify the value of a parameter J that controls the relative cost of
misclassifying seizure and non-seizure vectors. If the value of J is set so that the cost
of misclassifying a seizure vector far exceeds the cost of misclassifying a non-seizure
vector, then the SVM learning algorithm could produce a boundary that can cor-
 0.8 O% o o - "-o

0 4
0*

0.2 X i

0 0.2 0.4 0.6 0.8 1

Energy (0-16Hz)

Figure 3-16: Nonlinear decision boundary separating seizure (red) and non-seizure
(blue) feature vectors extracted from a single iEEG channel. The decision boundary
was determined using a radial basis kernel and the SVM learning algorithm.

rectly classify the first and second seizure vectors at the expense of classifying more
non-seizure vectors incorrectly.

The mapping function f(-) that establishes whether a feature vector falls within
the region of the feature space defined by seizure activity is expressed as shown in
equation 3.2 in the case of a linear boundary

f(X) = Seizure if WTX + f > 0 (3.2)

f(X) = Non-seizure if WTX +3P 0

Where the vector W, which is the normal to the separating hyperplane, and the
bias term 3 are parameters determined by the SVM learning algorithm [9]. In the
case of a nonlinear boundary determined using a radial basis kernel, the discriminant
function takes the form shown in equation 3.3. The coefficients aci, support-vectors X,
and bias term p/3are again determined by the SVM learning algorithm. The parameter
y, which controls how tightly the nonlinear boundary circumscribes a class, is user-
defined.
 Nsv
f(X) = Seizure if {- a exp(-yllX - Xjj12)} +± > 0 (3.3)
i=1
Nsv
f(X) = Non-seizure if { ai exp(-,y|X - XiI 2)} + < 0
i= 1

The number of support-vectors Nsv is partly governed by the complexity of the

classification task. As the similarity between an individual's seizure and non-seizure
activity increases more support-vectors are needed in order to define a more complex
decision boundary, and as a result the computational cost of equation 3.3 increases.

The support-vector machine algorithm is well-suited for the task of seizure onset

detection because its learning mechanism focuses on determining a decision boundary

that separates the boundary cases of each class; recall that the boundary cases of the
seizure class are associated with the onset of the seizure. Support-vector machines

have also been shown to perform well in high-dimensional classification tasks [27] and
in the setting where one class has a smaller number of training examples [1].

3.4 Patient-Specific Detector Architecture

The block diagram in Figure 3-17 illustrates the processing stages of the patient-
specific detector. The detector passes L-second epochs from each of N EEG channels
through a filterbank. In turn, the filterbank computes for each channel M features
that correspond to the energies within M frequency bands. The M features extracted
from each of the N channels are then concatenated to form a M x N element vector
XT that automatically captures the spectral and spatial relations between channels.

Next, the W feature vectors XT, XT-L,..., and XT-(w-1)L are concatenated to form

the feature vector XT. This vector, which contains W x Mx N elements, automatically

captures the time evolution of spectral and spatial relations among the input EEG
channels.

Finally, the feature vector Xr is assigned to the seizure or non-seizure class using a
 Spectral Features Spectral and Spatial Spectral, Spatial Support Vector Machine
Features and Temporal Classification
Features

B Test Feature Vector

S. . x2,1 S Training Seizure Feature Vectors
A Training Non-Seizure Feature Vectors

iT-L A

=XT AA
.. xlN A
EEG Ch N N 1 )L A
2,N A

. .. .. .. . .. .. .. .. .
...... ......
........

Dimension: M*N Dimension: W*M*N

Figure 3-17: Patient-specific seizure onset detector architecture.

two-class support-vector machine classifier with cost factor J, radial basis kernel with
parameter -y, and trade-off between classification margin and error C. The detector
declares a seizure after one feature vector XT is assigned to the seizure class.

Since seizure and non-seizure activity are generally stereotypical for a patient and
highly variable across patients (Chapter 2), the support-vector machine is trained
on training seizure and non-seizure vectors from a single patient. The training non-
seizure vectors are extracted from H hours of continuously recorded scalp EEG. The
seizure vectors are derived from the first S seconds following the onset of K training
seizures.

Figure 3-18 illustrates how successive non-seizure training vectors are extracted
from a non-seizure training record when L=2 and W=3. For these parameter values,
a six second window is shifted, one second at a time, across a non-seizure record. For
each positioning of the six second window within the non-seizure record, a training

non-seizure feature vector X is generated. More specifically, in Figure 3-18, the time

interval 0-6 seconds is broken up into three, non-overlapping 2 second epochs that are
used to generate the feature vectors X 2 , X 4 , and X 6 . These vectors are then grouped
to form the first training non-seizure feature vector X6 . Next, the time interval 1-7
seconds is broken up into three, non-overlapping 2 second epochs that are used to
generate the feature vectors X 3 , X 5 , and X 7. These vectors are then grouped to form
the second training non-seizure feature vector A7 . The third training non-seizure
feature vector Xs is formed by grouping the feature vectors X 4 , X 6 , and X 8 extracted
from the time interval 2-8 seconds. The training non-seizure vector 's overlaps the
training vector '6 in composition.

Z6 X8

X2 x 4 X6 X8

0 1 2 3 4 5 6 7 8 Time (Seconds)

X3 x 5 X7

X7

Figure 3-18: Generating training non-seizure feature vectors.

The process used to generate training seizure vectors is similar to that used to
generate training non-seizure vectors. Figure 3-19 illustrates how successive seizure
training vectors are extracted from a training seizure when L=2 and W=3. In Figure
3-19, seizure onset is at time 0 seconds. The first training seizure vector X2 is formed
by grouping the feature vectors extracted from the six second window spanning the
time interval between -4 and +2 seconds. The second training seizure vector A3 is
formed by grouping the feature vectors extracted from the six second window spanning
the time interval between -3 and +3 seconds. This process is repeated until one has
advanced S seconds into the seizure.

X2 X4

X-2 Xo X 2 X4

-4-3 -2 -1 0 1 2 3 4 Times (Seconds)

x- 1 X1 X3

Figure 3-19: Generating training seizure feature vectors.

Table 3.1 summarizes the parameters that control our patient-specific detection
algorithm.

3.5 Instantiating Detector Parameters

In this section we assign values to the detector parameters listed in Table 3.1. In
Chapter 5 we illustrate how perturbing the assigned value of some of these parameters
impacts seizure onset detection performance.

3.5.1 EEG Epoch Length: L

Most of the spectral energy of physiologic and pathophysiologic scalp EEG activity

falls within the 0.5-24 Hz frequency band [19]. In order to resolve the lowest end of

the this frequency range, the detector extracts spectral features from a L = 2 second
 Table 3.1: Patient-specific Detector Parameters

Parameter Description
L EEG epoch length
N Number of EEG channels
M Number of filters in filterbank
W Number of feature vectors that form XT
SSVM radial basis kernel parameter
J Relative cost of misclassifications
C Trade-off between classification margin and error
H # of hours used to extract training non-seizure vectors
S # of seconds into seizure used to extract training seizure vectors
K 0 of seizures used to extract training seizure vectors

epoch. Other seizure detection algorithms have used 2.5 second [46] and 6 second [37]
EEG epochs.

3.5.2 Number of EEG Channels: N

All available EEG channels are provided as inputs to the seizure detection algorithm.
The scalp EEG data set used to evaluate the algorithm, which is described in Chapter
4, contained N = 18 EEG channels. However, in [54], Shih shows that limiting the
detector presented in this chapter to using N = 5 channels does not significantly
impact seizure detection performance.

3.5.3 Number of filters: M

The detector uses an M = 8 uniform bandwidth filterbank. The fiterbank spans the
0-25 Hz frequency range using filters that are 3 Hz wide. In Chapter 5 we demonstrate
the impact of using a filterbank with M < 8 filters.

3.5.4 Number of feature vectors in XT: W

The detector groups W = 3 feature vectors to form the feature vector X-. With
W = 3, the detector examines the evolution of the EEG over a period of 6 seconds
when determining whether or not to declare a seizure. As will be shown in Chapter
5, using fewer vector to form XT increases false detections, and using more vectors
increases detector latency.

3.5.5 SVM Parameters: y, J, and C

The SVM used by the detector is trained using the SVMLight software package [28]
with a cost factor J = 1, RBF kernel parameter of 7 = 0.1, and trade-off between
classification margin and error C = 1. We showed in [55, 56] that these values, after

adjusting for SVM software package differences, work well on an independent data set
consisting of 139 seizures and 49 hours of non-seizure EEG recorded from 36 pediatric

epilepsy patients.

3.5.6 Training Parameters: H, K, and S

The detector should be trained on non-seizure vectors extracted from H > 24 hours
of non-seizure EEG so that it is exposed to awake, sleep, abnormal, and artifact-
contaminated EEG. The detector should be trained on as many seizures originating
from the same seizure focus; however, as shown in Chapter 5, performance gains
are marginal for K > 3 seizures. For any given seizure, the detector is trained on
seizure vectors extracted from the first S = 20 seconds of a seizure. Seizure onset

is marked by an electroencephalographer and corresponds to the onset of rhythmic

activity associated with a clinical seizure. In Chapter 5 we demonstrate the impact
of training on S < 20 seconds.
76
Chapter 4

Scalp EEG Data and Testing

Methodology

In this chapter we describe the scalp EEG data and testing methodology used to
estimate the latency, sensitivity, and specificity of the patient-specific seizure onset
detector described in Chapter 3.

4.1 Scalp EEG Data Set

The data set used to evaluate the performance of our patient specific detector consists
of continuous Scalp EEG recordings from 23 pediatric patients (age < 18) undergoing
medication withdrawal for epilepsy surgery evaluation at Children's Hospital Boston.
The EEG was sampled at 256 Hz and recorded using an 18-channel, 10-20 bipolar
montage. According to the 10-20 bipolar montage, EEG electrodes are arrayed on the
scalp as shown in Figure 2-1, and the EEG channels formed are those illustrated in
the EEG tracings discussed in Chapter 2. Overall, this 23 patient data set contained
844 hours of continuously recorded EEG and 163 seizures.

The scalp EEG data set is segmented into records. A record is typically one hour
long. Records that do not contain a seizure are called non-seizure records and those
that contain one or more seizures are called seizure records.
4.2 Performance Metrics

Three metrics will be used to characterize the performance of our seizure onset de-
tector.

1. Electrographic Seizure Onset Detection Latency EOLaten,,,. The electro-

graphic onset of a seizure refers to the onset of scalp EEG changes associated
with a seizure. The clinical onset of a seizure refers to the onset of its physical
or cognitive symptoms. In scalp EEG, the electrographic onset of a seizure may
or may not precede its clinical onset; in contrast, in iEEG, the electrographic
onset of a seizure tends to precede its clinical onset. EOLate,, refers to the
delay between electrographic onset and detector recognition of seizure activity.
By definition EOiten > 0.

2. Sensitivity S refers to the percentage of test seizures identified by a detector.

3. False Alarms Per Hour FA refers to the number of times, over the course
of an hour, that a detector declares the onset of seizure activity in absence of
an actual seizure.

4.3 Performance Metric Measurement

To estimate our detector's performance on data recorded from a patient, we use a

leave-one-record-out cross-validation scheme. For the purposes of this discussion, Let
NNS denote the number of non-seizure records and let Ns denote the number of
seizure records.

Measuring EOLateny and S

The detector is trained on the NNS non-seizure records of a patient (median NNS =
33) as well as Ns - 1 records containing a seizure (median Ns = 5). The detector
is then tasked with detecting seizures in the mth seizure record; the record that was
withheld from the training set. This process is repeated Ns times so that each of
the Ns seizure records is tested once; a seizure record is never simultaneously in the
training and testing sets.
Let Sm E {0, 1} be a binary variable that denotes whether the detector noted the

mth -seizure, and let EOLatency,m denote the latency with which the detector notes
that seizure's electrographic onset. Let FAs,m be the number of false alarms declared

while processing the mth seizure record. Finally, let K denote the total number of
detected seizures. Equation (4.1) shows how these quantities are combined into an
estimate of the detector's performance for a given patient.

1 NS
S E Sm (4.1)
m=1
EOLatene = m * EOLatency,m

Measuring FA

A detector is trained on the Ns seizure records of a patient as well as NNs - 1

non-seizure records. The detector is then tasked with processing the nt h non-seizure
record; the record that was withheld from the training set. This process is repeated
NNS times so that each of the NNs non-seizure records is tested once; a non-seizure
record is never simultaneously in the training and testing sets. Let FANs,n be the
number of false alarms declared while processing the nth non-seizure record. Equation
(4.2) is used to estimate the detector's false alarm rate.

NNS Ns
FA = NN+ Ns * ( FAN, + FAs,m) (4.2)
n=l m=l
80
Chapter 5

Performance

In this chapter we evaluate the performance of the detector described in Chapter

3 using the scalp EEG data and testing methodology described in Chapter 4. We
also analyze how varying certain detector parameters impacts detector performance.
Finally, we compare the performance of our algorithm to that of a state-of-the-art
patient non-specific seizure detector.

5.1 Patient-Specific Detector Performance

5.1.1 Latency

Figure 5-1 illustrates the percentage of test seizures detected within a specified la-
tency. As can be seen, 50% of the 163 test seizures are detected within 3 seconds,
71% within 5 seconds, and 91% within 10 seconds. The mean latency with which the
detector declared the onset of a test seizure is 4.6 seconds.
Figure 5-2 illustrates the latency with which the detector declared the onset of
each seizure for each of the 23 subjects. A seizure is represented by a blue dot, and
the height of a dot represents the latency associated with a seizure detection. If more
than one seizure was detected with the same latency, then the numeral next to a dot
indicates the number of seizures it represents. For most patients, the majority of
seizures are detected within 5 seconds.
 0.9 91% Detected within 10 seconds

0.8

E 0.7 71% Detected within 5 seconds

0 0.6
.-

- 0.5 50% Detected within 3 seconds

! 0.4

0.3

0.2

0.1
0 2 4 6 8 10 12 14 16 18 20
Detection Latency (Seconds)

Figure 5-1: Percentage of 163 test seizures detected within a specified latency. The
detector notes the onset of 50% of all test seizures within 3 seconds.

For Patient 15 one of the seizures is detected with a 55 second delay. The reason
for the long detection latency is that the EEG at the onset of this seizure differed

in spectral and spatial characteristics from training seizures. Figure 5-3 illustrates a

typical training seizure. The seizure begins following 272 seconds with a theta band
rhythm that is most prominently seen on the channel T7 - P7. Figure 5-4 illustrates
the seizure that the algorithm failed to detect promptly. This seizure, which begins

at 876 seconds, involves a train of spikes on the channel P7 - 01.

5.1.2 Sensitivity

Figure 5-5 shows the sensitivity with which the patient-specific algorithm detects
the test seizures of each of the 23 subjects. Overall, 96% of 163 test seizures were
detected.

5.1.3 Specificity

Figure 5-6 shows the number of false detections declared by the patient-specific de-

tector in a period of 24 hours for each of the 23 subjects. For most patients (18 of
 Seizure Onset Detection Delay
60

50 - - ---- - --

50 ------ -

40

20 ------------
-
20

is ------
-- - - -- C-~--
---
----------

10 0* ------ ----------- ----------

--

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Patient Number

Figure 5-2: Latency with which the detector notes the onset of seizures for each of
the 23 test subjects. Each dot represents a seizure. A numeral is placed next to dots
that represent more than one seizure. For most patients, the majority of seizures are
detected within 5 seconds.

F7-T7
TT-P7L

FP1-
F3-C3 ....;I fa4 MF
40"I1--A
I

P3-0

C4-P

FP2-

F8-T8

P8-
FZ-
CZ-
264 266 268 270 272 274 276 278 280 282 284

Figure 5-3: Example of a seizure within the scalp EEG of Patient 15. The seizure,
which begins at 272 seconds, consists of a theta band rhythm that is most prominently
seen on the channel T7 - P7.

83
 FP1-F

C3-P -F t r

P8-0T8
FZ-C4 - -- I

CZ-P
868 870 872 874 876 878 880 882 884 886 888

Figure 5-4: Example of another type of seizure observed within the scalp EEG of
Patient 15. This seizure, which begins at 876 seconds, consists of a train of spikes
on the channel P7 - 01. The detector fails to detect the onset of this seizure since
its spectral and spatial characteristics differ from those of training seizures. Training
seizures resemble the seizure shown in Figure 5-3.
U Detected Seizures MTotal Number Of Seizures
26 . .. ... ..

24
24
-
--- -- --... - . -.. . ..-------.

22

20

18
--
---
16---

S14

2 101 ---- ---

2..
1
2 . . . ...
2 3 4
..
.....
5
... ..
.....
6
.....
.....
7
.....
...
.
8 9 10
I 7 11

Patient Number
12 13 14 15 16 17 18 19 20 21 22 23

Figure 5-5: Sensitivity of the patient-specific seizure detector. Red bars show the
number of test seizures available for each subject, and black bars show the number
of test seizures recognized by the detector. Overall, the detector recognized 96% of
163 test seizures.
 23), the detector declares between 0 and 5 False detections per 24 hours.

22

20

18

16

E 8
14
;10
i _ _ _ _ __ _ _

2
6

0 ~I
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Patient Number

Figure 5-6: Specificity of the patient-specific seizure detector. For most patients (18
of 23), the detector declares between 0 and 5 false detections per 24 hours.

For Patient 13 the algorithm declares 20 false detections per 24 hour period. The
reason for the high false alarm rate is the presence of short bursts of rhythmic activity
that match the seizure onset signature both in spectral and spatial character. Figure
5-7 illustrates the onset of a typical seizure recorded from Patient 13. The seizure
begins following 640 seconds with a delta band rhythm that is most prominent on the
channel FP1 - F7. Figure 5-8 illustrates the burst of rhythmic activity that resulted
in a false detection. The burst, which occurs between 233-235 seconds, also consists
of a delta band rhythm that is most prominent on the channel FP1 - F7. The main
feature that distinguishes the seizure from the burst is the seizure's temporal extent.
To improve the specificity of the detection algorithm for Patient 13, one could declare
a seizure only when activity suspected of being a seizure persists for a duration of

time that exceeds the average length of the bursts of rhythmic activity. The cost of

such a modification will be an increase in detection latency.

For the two patients with the highest number of false detections, patients 12 and
 FP1
F7-TT
FPig E 5-7 o a seizur withi th s EEG of Patient 13. The seiure

C3-P
P3-0
that is most prominenton the cn
rhyth
FP2- FP F Since
- the burt r

F8-T8

634 636 638 640 642 44 646 648 650 652 654

Figure 5-7: Example of a seizure within the scalp EEG of Patient 13. The seizure,
which begins at 640 seconds, consists of a delta band rhythm that is most prominent
on the channel FP1 - F7.

TT-

P3-0
FP

F8 T 8 ......
. I I.

P8-0-

223 225 227 229 231 233 235 237 239 241 243

Figure 5-8: Example of a short burst of rhythmic activity within the scalp EEG of
Patient 13. The burst, which occurs between 233-235 seconds, consists of a delta band
rhythm that is most prominent on the channel FP1 - F7. Since the burst resembles
the seizure in Figure 5-7, the detector declares a false detection following the onset
of the burst.
13, we note that false detections tend to occur in a small fraction of the 1 hour records
processed by the detector. Figure 5-9 illustrates that 75% of the test records (24 of

33 hours) resulted in 0 false detections in the case of Patient 13. For Patient 12, 68%
of the test records (11 of 16 hours) resulted in 0 false detections.

--- Patient 13
0.9 ---- Patient 12
0.8

0.7

0.6

,, 0.5
o.4
- 0.4

g 0.3
I.
0.2

0.1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Percent of Test Records

Figure 5-9: False detections declared by the detector are not uniformly distributed
across test records. For Patient 13, 75% of the test records resulted in 0 false detec-
tions. For Patient 12, 68% of the test records resulted in 0 false detections.

5.2 Varying Patient-Specific Detector Parameters

5.2.1 Varying the Number of Filters: M

On the scalp EEG data set discussed in section 4.1 we profiled the performance of our
patient-specific detector as a function of uniformly splitting the bandwidth 0-25 Hz
using 2, 4, and 8 equal bandwidth filters. In these experiments the number of filters
M used to generate the feature vector XT varies, but the number of feature vectors

that are grouped to form XT is fixed at W=3.

Increasing the number of filters within the filterbank does not greatly impact the

detection latency or the sensitivity of the detector. The average, across 23 subjects,
latency of detectors with 2, 4, and 8 filter filterbanks was 4.4, 4.7, and 4.8 seconds

respectively, and all detectors had a sensitivity of 96%. Increasing the number of
filters appears to improve the specificity of a detector as shown in Figure 5-10. The

average specificity of a detector that uses a two-filter filterbank is greater than 8 false
detections per 24 hours, and that of a detector with an eight-filter filterbank is less
than 4 false detections per 24 hours.

4
Number of Fhers

Figure 5-10: The false detection rate of the detector decreases as number of filters
used to construct XT is increased from 2 to 8. A two-filter filterbank yields a false
detection rate of 8 false detections per 24 hours. An eight-filter filterbank results in
a false detection rate smaller than 4 false detections per 24 hours.

5.2.2 Varying the Number of Feature Vectors in XT: W

On the scalp EEG data set discussed in section 4.1 we profiled the performance of
our patient-specific detector as the number feature vectors W used to form XT is

changed. In these experiments a M = 8 filter filterbank was used to generate the

feature vectors XT, XT-L,..., XT-(W-1)L that are embedded within Xr.

The sensitivity of the patient-specific detector remained 96% as the number of

 feature vectors within XT was increased from 2 through 4. However, as W is increased
the false detection rate of the detector decreases while its latency increases as shown
in Figure 5-11. In order to detect seizures with an average latency shorter than 5
seconds one should set W = 3.

3.5 4 4.5 5 5.5

Latency(Seconds)

Figure 5-11: Increasing the number of feature vectors within Xr decreases the detec-
tor's false detection rate and increases its detection delay. To detect seizures with an
average latency shorter than 5 seconds one should set W = 3.

5.2.3 Varying the Number of Training Seizures: K

The performance of the detector improves as more seizures are included in the training
set. Figure 5-12 illustrates how the average detection latency and miss rate decrease
with an increasing number of training seizures for five randomly selected patients.
With a single training seizure, the detector has a latency greater than 7 seconds and
misses more than 45% of the test seizures. With three training seizures, the detector

has a latency close to 4 seconds and misses less than 5% of test seizures. The addition

of a fourth seizure improves the detector's latency to 3.5 seconds but does have a great

impact on the detector's miss rate.

 -4-Latency -U-MISS Rate

8 ------- 0.5

0.45

OA
0.4
6
0.35

4 0.252

3 -------------- - o---
0.15

0.1

0.05

0 - . ..........
....... .....
....
--.. .. .... 0
..

1 2 3 4
Numberof TrainingSeizures

Figure 5-12: Increasing the number of training seizures decreases the detector's miss
rate and its detection delay. Including more than three training seizures results in a
marginal improvement in detector's miss rate.

5.2.4 Varying Training Time into a Seizure: S

In this section we examine how the performance of the detector changes when training
on seizure feature vectors derived from the first S = 6, 12, 18, or 24 seconds of training
seizures. Figure 5-13 shows that the detector's average latency does not improve
significantly for S > 12 seconds. Moreover, the detector's sensitivity exceeds 90%
for S > 18 seconds as shown in Figure 5-14. Finally, Figure 5-15 shows that the
detector's false detection rate increases as the value of S is increased.

One explanation for the increase in false detections with increasing S assumes
that the decision boundary developed by the SVM learning algorithm encloses a
single, contigous region of the feature space as illustrated in Figure 3-16. With this
assumption, increasing S forces the SVM to develop a wider decision region in order to

enclose feature vectors from the early and late portions of a seizure. Such a boundary

is more likely to result in false detections since it does not tightly circumbscribe the
training seizure vectors.
 3.5

2.5

1.5

0.5

0
6 12 18 24
Seconds Fol Seizurm
Onset

Figure 5-13: Including seizure vectors derived from the first S > 12 seconds of each
training seizure does not improve seizure detection delay significantly. For S > 12
seconds the detector's mean latency is less than 3 seconds.

1.2

S0.1

.2

0.2

0
6 12 18 24
Secmn Following
Seure Onset

Figure 5-14: Including seizure vectors derived from the first S > 18 seconds of each
training seizure does not improve seizure detection rate significantly. For S > 18
seconds the detector recognizes more than 95% of all test seizures.
 I

0
6 12 18 24
Seconds FollowinX
Seizure Onset

Figure 5-15: Increasing S increases the detector's false detection rate.

5.3 Patient-Specific and Non-specific Seizure De-

tection

In this section we compare the performance of our patient-specific detector to the

performance of the Reveal algorithm, a patient non-specific detector. Recall that Re-
veal examines one EEG channel at a time and then employs hand-coded and neural
network rules to determine whether features derived from a time-frequency decompo-
sition of a channel are consistent with a seizure event. The thresholds for some of the
neural network rules are determined using both archetypal seizures from individuals
with epilepsy and background EEG from individuals without epilepsy.

5.3.1 Performance Comparison

We evaluated the performance of the Reveal algorithm under a sensitive and a specific
setting. Under the sensitive setting, the Reveal algorithm declares a seizure whenever
5 seconds of EEG resembles training seizures with a score of 0.9 out of 1. Under the
specific setting, a seizure is declared whenever 20 seconds of EEG resembles training
seizures with a score of 0.9 out of 1.
Under the sensitive setting, the Reveal algorithm detected 74% of 152 test seizures
while our patient-specific method detected 96% of the test seizures. Figure 5-16
shows, for each patient, the number of seizures detected by our method and the
Reveal algorithm.

Performance Comparison: Sensitivity

a RevealSensitive Setting U Patient Specific Detector MTotal Number Of Seizures
22

20

16

E
E 14

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Patient Number

Figure 5-16: Comparison of the number of test seizures recognized by the Reveal
(sensitive setting) and patient-specific algorithms. Black bars show the number of
test seizures available for each subject, red bars show the number of test seizures rec-
ognized by the patient-specific algorithm, and blue bars show the number of seizures
recognized by the Reveal algorithm. The Reveal algorithm detected 74% of 152 test
seizures and our patient-specific method detected 96% of all test seizures.

The specificity of the Reveal algorithm was poor as illustrated by the number
of false detections declared in a 24 hour period (Figure 5-17). For some patients
the Reveal algorithm declared an excess of 100 false detections per 24 hour period.
Figure 5-18 expands the y-axis of Figure 5-17 so that the false detection rates of the

patient-specific detector are visible.

Under the specific setting, the sensitivity of the Reveal algorithm decreased to

61%. A comparison, for each patient, of the number of seizures detected by our
 Performance Comparison: False Alarms Per 24 Hours
a Reveal Sensitive Setting a Patient Specific Detector

350

0 300

200

150

100

0 ..

1 2 3 4 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Patient Number

Figure 5-17: Comparison of the number of false detections declared by the Reveal
(sensitive setting) and patient-specific algorithms. For some patients the Reveal al-
gorithm declared an excess of 100 false detections per 24 hour period.

Performance Comparison: False Alarms Per 24 Hours

a Reveal Sensitive Setting a Patient Specific Detector

22

20

18
16

14

12
10

1 2 3 4 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Patient Number

Figure 5-18: Comparison of the number of false detections declared by the Reveal
(sensitive setting) and patient-specific algorithms. Expanded y-axis.
method and the Reveal algorithm is shown in Figure 5-19. As expected, the specificity
of Reveal improves under this setting but it remains worse than that of the patient-
specific detector as shown in Figures 5-20 and 5-21.

Performance Comparison: Sensitivity

U Reveal Specific Setting a Patient Specific Detector 0 Total Number Of Seizures
22

20
is - --------- ----- -- ------------------------- --- -- ------ ----

16 -

S14

12
2

% 10
E

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Patient Number

Figure 5-19: Comparison of the number of test seizures recognized by the Reveal
algorithm (specific setting) and our patient-specific algorithm. Black bars show the
number of test seizures available for each subject, red bars show the number of test
seizures recognized by the patient-specific algorithm, and blue bars show the number
of seizures recognized by the Reveal algorithm. The Reveal algorithm detected 61%
of 152 test seizures and our patient-specific method detected 96% of all test seizures.

5.4 Case Studies

The following examples illustrate why a patient-specific approach can yield better
seizure detection performance.

5.4.1 Latency

Figure 5-22 illustrates a seizure recorded within the scalp EEG of Patient G. The

onset of the seizure occurs at 701 seconds and consists of beta band activity that is
 Performance Comparison: False Alarms Per 24 Hours
a Reveal Specific Setting U Patient Specific Detector

200 --

150

100 ------

1 2
1 2 3 4 5 6 7 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
34567
Patient Number

Figure 5-20: Comparison of the number of false detections declared by the Reveal
(specific setting) and patient-specific algorithms. For some patients the Reveal algo-
rithm declared an excess of 100 false detections per 24 hour period.

Performance Comparison: False Alarms Per 24 Hours

0 Reveal Specific Setting U Patient Specific Detector

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Patient Number

Figure 5-21: Comparison of the number of false detections declared by the Reveal
(specific setting) and patient-specific algorithms. Expanded y-axis.
 most prominent on the channels {F4 - C4, C4 - P4, F8 - T8, T8 - P8}. Twenty-nine
seconds into the seizure, large amplitude theta band activity can be seen on all EEG
channels as shown in Figure 5-23.

FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-F8
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
699 701 703 705

Figure 5-22: Seizure recorded within the scalp EEG of Patient G. The seizure, which
begins at 701 seconds, consists of beta band activity that is most prominent on the
channels {F4 - C4, C4 - P4, F8 - T8, T8 - P8}.

In this case, our patient-specific detector demonstrated improved detection la-

tency relative to the Reveal algorithm. The patient-specific detector, on average,
recognized that a seizure was ongoing within 7 seconds of its electrograpihc onset.
On the other hand, the Reveal algorithm detected the presence of seizure activity
21.5 seconds following seizure onset. The Reveal algorithm could not detect the sub-
tle EEG change associated with seizure onset, but could detect the high-amplitude,
generalized activity that emerged towards the end of the seizure.

5.4.2 Sensitivity

Figure 5-24 illustrates a seizure recorded within the scalp EEG of Patient H. The
seizure begins at 977 seconds with rapid eye-blinking that manifests as large, down-
ward deflections on the channels FP1 - F3, FP2 - F4, FP1 - F7, FP2- F8. Later,
 FP1-F7
F7-T7
T7-P7
P7-01
FP1-F3
F3-C3
C3-P3
P3-01
FP2-F4
F4-C4
C4-P4
P4-02
FP2-FB
F8-T8
T8-P8
P8-02
FZ-CZ
CZ-PZ
725 77 729 731

Figure 5-23: Large amplitude theta band activity can be seen on all EEG channels
twenty-nine seconds after the onset of the seizure shown in Figure 5-22.

at 983 seconds, the eye-blinks are accompanied by theta band activity on the channels
{T4 - T6, T6 - 02}.

In this case, the Reveal algorithm failed to detect any of the test seizures. A

possible explanation for Reveal's poor sensitivity is the overlap of eye-blinks on the
frontal channels (channels with the letter F) with seizure discharges on the temporal
channels (channels with the letter T). Patient non-specific algorithms, such as the
Reveal algorithm, are generally designed to classify EEG epochs containing rapid
eye-blinks as corrupt and unlikely to contain a seizure. In contrast, our patient-

specific detector could detect these seizures because it learned that the key feature
that differentiates seizure-associated eye-blinks from others is the presence of seizure
discharges on the temporal channels.

5.4.3 Specificity

Figure 5-26 illustrates a seizure recorded within the scalp EEG of Patient J. The

onset of this seizure invovles a generalized spike at 12231 seconds that is followed

by generalized rhythmic activity. Both the Reveal and patient-specific algorithms

 Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5

Fp2-F8
F8-T4
T4-T6
T6-02Jp
977 979 981 983 985 987 989 991 993

Figure 5-24: Seizure recorded within the scalp EEG of Patient H. The seizure, which
begins at 977 seconds, consists of rapid eye-blinking that is later accompanied by
theta band activity on the channels {T4 - T6, T6 - 02} at 983 seconds.

were successful in detecting the onset of seizures recorded from Patient J. However,
the Reveal algorithm exhibited poor specificity because it declared false detections
whenever the rhythmic activity illustrated in Figure 5-26 occurred. This rhythmic
activity is commonly observed within the awake EEG of Patient J. The patient-specific
detector correctly classified the rhythmic activity as belonging to the non-seizure class
because it had been labeled as such in the training set.
Figure 5-25: Seizure recorded within the scalp EEG of Patient J. The onset of this
seizure invovles a generalized spike at 12231 seconds that is followed by generalized
rhythmic activity.

FP1
F7-T
T7-P
P7-O O ll IIIP

F3-

FP2
F4-
C4- mill01-04-0) 0 "M I
1_710, I

Figure 5-26: Rhythmic activity commonly observed within the awake scalp EEG of
Patient J.

100
Chapter 6

Seizure Onset Detection Using

Physiologic Signal Fusion

Using the scalp electroencephalogram (scalp EEG) to detect seizure onsets that are
not associated with rhythmic EEG activity is challenging. In this chapter we illus-

trate how supplementing information extracted from the scalp EEG with information
extracted from a second physiologic signal can improve the detection of these types
of seizures.

6.1 Why Use a Second Physiologic Signal?

As discussed in Chapter 2, seizures involve the hyperactivity and hypersynchrony of

a population of neurons. At the level of the scalp EEG, this coherent neural firing
gives rise to rhythmic activity with a dominant frequency between 0-25 Hz. However,
when the underlying neural hypersynchrony involves a neural network deep within
the brain, the earliest scalp EEG changes may not reflect the underlying neuronal
hypersynchrony, but physical consequences of the seizure such as rapid eye-blinks or
muscle contractions. Once the entrained nerual mass is large enough, rhythmic ac-

tivity reflective of neuronal hypersynchrony becomes manifest within the scalp EEG.

Detecting seizures of this kind using the scalp EEG is challenging since the EEG

changes observed following seizure onset are also seen routinely during non-seizure

101
states. In fact, classic seizure detection methods insert processing stages to remove
such activity [34, 48] because detecting it would result in poor specificity.
In order to detect these types of seizures a detector requires information beyond
that within the scalp EEG to ascertain whether or not a seizure is taking place.
The additional information can be derived using a second physiologic signal whose
dynamics are influenced by the seizure. The second physiologic signal and the scalp
EEG will complement each other and improve seizure onset detection if the changes,
in each of these signals, that suggest the onset of a seizure rarely coincide during non-
seizure states and often coincide at the time of an actual seizure. Patient-specificity
remains essential to the success of this approach since the manner with which the
scalp EEG and the secondary signal change during seizure and non-seizures states
varies across patients.

6.2 The Electrocardiogram as a Second Signal

A variety of physiologic signals may be used to supplement the scalp EEG in order to
improve seizure detection performance. For example, seizures resulting in repetitive
motor activity may become readily detectable if scalp EEG data is supplemented with
accelerometer sensor data [40]. For other types of seizures, especially those originating
within or spreading to the temporal lobes, seizures are associated with electrocardio-
graphic changes. The most common ECG change associated with seizures is a heart
rate acceleration (tachycardia) [33].

6.3 Patient-Specific, EEG-ECG-based Seizure De-

tection

As was the case in Chapter 3, a classifier will be used to determine whether an

observed feature vector is representative of an individual's seizure or a non-seizure
state; however, in this chapter, the feature vector will contain information extracted
from both the EEG and ECG.

102
 Spectral EEG Features Spectral and Spatial Spectral, Spatial Support Vector Machine
EEG Features and Temporal Classification
EEG/ECG Features

x,1
EEG Ch 1 x B Test FeatureVector
21 * Training. Seizure Feature Vectors
X Training Non-Seizure Feature Vectors

xlN

EEG Ch N ' xiN L AA 6 I

2, N XEC AA

ECG Features

ECG
Mean Heart Rate

Heart Rate C = XT,ECG

Change 2,ECG
:................................

Figure 6-1: Block diagram of patient-specific seizure onset detector that combines
features extracted from the EEG and ECG signals.

The block diagram in Figure 6-1 illustrates the architecture of the combined EEG

and ECG patient-specific seizure onset detector. The detector passes a L second

epoch from N = 16 EEG channels and a lead-II ECG channel through a feature
extractor. The feature extractor uses a filterbank with M = 8 filters to compute for
each EEG channel the energy in eight overlapping frequency bands; the frequency
bands are each 3 Hz wide and span 0-25Hz. The 8 features extracted from each of
the 16 channels are then concatenated to form a 128 dimensional feature vector XT
that automatically captures the spectral and spatial correlations between channels.

The feature extractor uses the ECG channel to extract a feature vector XT,ECG
that is composed of two features. The first feature, X1,ECG, corresponds to the mean
heart rate within the L second epoch. The second feature, X2,ECG, corresponds to
the difference between the instantaneous heart rate measured at the start and end

of the L second epoch. Outside the context of a seizure, X2,ECG = 0 as successive,

instantaneous heart rate measurements within the L second epoch will oscillate about

a mean heart rate. In contrast, for patients experiencing seizure-associated tachycar-

103
dia, X2,ECG > 0 since the instantaneous heart rate at the end of the epoch will exceed
that at the start of the epoch.
Next, W = 3 EEG feature vectors XT, XT-L, and XT-2L as well as the ECG
feature vector XT,ECG are concatenated to form a single feature vector XT. This
feature vector encodes the time evolution of spectral and spatial properties of the
scalp EEG as well as heart rate and heart rate change information extracted from the
ECG. This encoding is automatically generated without user intervention. Finally,
the feature vector XT is assigned to the seizure or non-seizure state using a two-class
support vector machine (Chapter 3).

6.4 Case Studies

In this section we present two case studies that illustrate how a detector that combines
information extracted from the scalp EEG and surface ECG can perform better than
a detector that only uses information extracted from the scalp EEG.

6.4.1 Data

The data set used to assess our detection methodology was gathered from two pa-
tients. From the first patient, an adult woman with refractory complex partial
seizures, 66 hours of continuous EEG-ECG that includes 10 seizures were collected.
From the second patient, an adult male with a childhood-onset epilepsy syndrome
known Lennox-Gastaut, 34 hours of continuous EEG-ECG that includes 3 seizures
were collected.

6.4.2 Case 1

Figure 6-2 shows an EEG tracing of a typical seizure recorded from the first subject.
The onset of the seizure, at 1486 seconds, involves rapid eye-blinking (eye flutter),
which manifests in the EEG as high-amplitude deflections on the channels {FP1 -
F3, FP2 - F4, FP1 - F7, FP2 - F8}.

104
 Eye Flutter

Fpl -F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F
FT-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

ECG TM

Lij Time (Seconds)

I
HR Acceleration Theta Waves
Figure 6-2: Example of a seizure within the scalp EEG of Case 1. The seizure,
which begins at 1486 seconds, involves rapid eye-blinking that results in downward
deflections on frontal EEG channels (e.g. {FP1 - F3, FP2- F4}). Coincident with
the onset of rapid eye-blinking, the patient's heart rate accelerates as shown in Figure
6-3. Later, at 1492 seconds, a 3-4 Hz theta wave appears on the EEG channel T4-T6.

With the onset of eye-flutter, the patient's heart rate accelerates as can be seen on
the bottom channel in Figure 6-2 and in the heart rate profile in Figure 6-3. Finally,
at 1492 seconds, 3-4 Hz theta waves appear on the EEG channel T4 - T6, while the

patient's heart rate remains elevated. Electrographic signs of the seizure end around
1510 seconds, but clinical symptoms of the seizure persist for another 2-3 minutes.

Performance Comparison

To evaluate the utility of combining EEG and ECG information in this case, we
compared the performance of two detectors. One detector classifies a feature vector
synthesized using EEG and ECG as in Figure 6-1, and the other classifies a feature
vector synthesized using EEG features as in Figure 3-17 of Chapter 3. Both detectors

are trained on the S = 20 seconds following the onset of eye-flutter within training

seizures, and both detectors process L = 5 second epochs. The SVM used within

the detector that combined EEG and ECG features was trained with 7 = 0.007 and

105
 140 r

130

120

110

a.
U) 100 Seizure Onset

90

80

70
1460 1465 1470 1475 1480 1485 1490 1495 1500
Time (Seconds)

Figure 6-3: Seizure onset, at 1486 seconds, is associated with an acceleration of the
patient's heart rate.

J = 5, while the SVM used within the detector that only used EEG was trained
with y = 0.003 and J = 5. The SVM parameters were chosen to bias each detector
towards recognizing seizure onset with a short delay.

Figure 6-4 illustrates the mean, minimum, and maximum latency with which
each detector declares the ten test seizures, and Figure 6-5 shows the false alarm rate
exhibited by each detector while analyzing the 66 hours of non-seizure test data.

In this case, the detector that combined EEG and ECG information detected
seizures with a mean detection latency (2.7 seconds) that is shorter than that of the
detector which relied solely on EEG information (4.2 seconds). Furthermore, the
detector that fused EEG and ECG information had a false detection rate (less than

5 false detections per 24 hours) that is lower than that of the detector which relied
on EEG information alone (more than 10 false detections per 24 hours).

6.4.3 Case 2

Figure 6-6 shows an EEG tracing of a typical seizure recorded from the second subject.

The onset of the seizure, at 56 seconds, consists of an electrodecrement involving all

106
 EEG+ ECG

Figure 6-4: Comparison of the minimum, maximum, and mean detection delays of two
detectors. One detector classifies a feature vector composed solely of EEG features,
while the other uses a feature vector that combines EEG and ECG features. The
detector that fuses features extracted from the EEG and ECG signals has a shorter
mean detection delay.

EEG+ ECG

Figure 6-5: Comparison of the false detection rates of two detectors. One detector
classifies a feature vector composed solely of EEG features, while the other uses a
feature vector that combines EEG and ECG features. The detector that fuses features
extracted from the EEG and ECG signals has a smaller false detection rate.

107
EEG channels and lasting for 12 seconds. Next, at 68 seconds, 1-2 Hz generalized,
rhythmic activity develops and persists for another 30 seconds. Clinically, throughout
the period of the electrodecrement the patient experiences tonic contraction of major
muscle groups.

Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

ECG Ai l

46 48 50 52 54 56 60 62 64 66 6 70
Time (Seconds)

Figure 6-6: Example of a seizure within the scalp EEG of Case 2. The seizure, which
begins at 56 seconds, involves a 12 second period of low-amplitude EEG activity
across most EEG channels. At the same time, the patient's heart rate accelerates
as shown in Figure 6-7. Later, at 68 seconds, 1-2 Hz generalized, rhythmic activity
develops.

With the onset of the electrodecrement, the patient's heart rate accelerates as can
be seen on the bottom channel in Figure 6-6 and in the heart rate profile in Figure
6-7. Throughout most of the period of generalized, rhythmic activity the patient's
heart rate remains elevated.

Performance Comparison

To evaluate the utility of combining EEG and ECG information in this case, we
compared the performance of two detectors. One detector classifies a feature vector

synthesized using EEG and ECG as in Figure 6-1, and the other classifies a feature

108
 ~110

S100
I. Seizure Onset

90 -

80

70

60
25 30 35 40 45 50 55 60 65
Times (Seconds)

Figure 6-7: Seizure onset, at 56 seconds, is associated with an acceleration of the

patient's heart rate.

vector synthesized using EEG features as in Figure 3-17 of Chapter 3. Both detectors
are trained on the S = 20 seconds following the onset of the electrodecrement within
training seizures, and both detectors process L = 2 second epochs. Similar to Case 1,
the SVM used within the detector that combined EEG and ECG features was trained

with y = 0.007 and J = 5, while the SVM used within the detector that only used
EEG was trained with y = 0.003 and J = 5.

Figure 6-8 illustrates the mean, minimum, and maximum latency with which each
detector declares the 3 test seizures. Figure 6-9 shows the false alarm rate exhibited
by each detector while analyzing the 34 hours of non-seizure test data.

In this case, the detector that combined EEG and ECG information detected
seizures with a mean detection latency (6 seconds) that is comparable to that of the

detector which relied solely on EEG information (5 seconds). However, the detector

that fused EEG and ECG information had a false detection rate (4 false detections
per 24 hours) that is lower than that of the detector which relied on EEG information

alone (more than 10 false detections per 24 hours).

109
 i
, 4

3-

1-

a-.
ECG+ ECG

Figure 6-8: Comparison of the minimum, maximum, and mean detection delays of two
detectors. One detector classifies a feature vector composed solely of EEG features,
while the other uses a feature vector that combines EEG and ECG features. The two
detectors have comparable seizure detection delays.

ECG+ ECG

Figure 6-9: Comparison of the false detection rates of two detectors. One detector
classifies a feature vector composed solely of EEG features, while the other uses a
feature vector that combines EEG and ECG features. The detector that fuses features
extracted from the EEG and ECG signals has a smaller false detection rate.

110
6.5 Importance of Patient-Specificity

We evaluated the performance of the Reveal algorithm (section 1.4.1) on Cases 1

and 2 to demonstrate the importance of both physiologic signal fusion and patient-
specificity to detecting the class of seizures discussed in this chapter. The Reveal
algorithm is only capable of processing EEG signals and is patient non-specific. When
tested on Case 1, the Reveal algorithm detected 1 of 10 seizures. When tested on
Case 2, the Reveal algorithm missed the electrodecrement in 2 out of 3 test seizures,
and instead detected the generalized 1-2 Hz activity that develops near the end of
the test seizures.

111
112
Chapter 7

Seizure-Triggered Vagus Nerve

Stimulation

In this chapter we illustrate the feasibility of applying the seizure detection method-
ology developed in Chapters 3 and 6 to the initiation of a delay-sensitive therapeutic

application. More specifically, the delivery of vagus nerve stimulation following seizure
onset.

7.1 Vagus Nerve Stimulation

For most of the history of epilepsy, pharmacologic, surgical, and diet-based treatments

were the only therapeutic options available to patients with refractory epilepsy. In
1997, the United States Food and Drug Administration approved the use of Vagus
Nerve Stimulation (VNS) as an adjunctive therapy for the treatment of refractory
seizures; the Vagus Nerve Stimulator became the first implantable device for the
treatment of epilepsy [3, 50, 51]. The VNS pulse generator is implanted in an infra-
clavicular, subcutaneous pocket and delivers stimuli to the midcervical portion of the
left vagus nerve.

Vagus Nerve Stimulation therapy is delivered in two modes. In automatic mode,

the implanted pulse-generator automatically delivers vagus nerve stimuli at pro-

grammed intervals. In on-demand mode, the patient or their caregiver initiates vagus

113
nerve stimulation in response to symptoms of a seizure. Initiating on-demand stimu-
lation requires holding a permanent magnet for 1-2 seconds over the implanted pulse

generator.

Several clinical trials have been conducted to assess the therapeutic efficacy of
automatic mode vagus nerve stimulation. The E03 and E05 trials demonstrated
that automatic mode VNS reduces seizure frequency by more than 50% in more

than 20% of patients within 3 months of device implantation [21, 23]. The XE5
trial demonstrated that the efficacy of automatic vagus nerve stimulation is both
long-lasting and improves significantly with time [11]. Based on these and other
studies, vagus nerve stimulation was deemed a safe and effective adjunctive therapy
for intractable epilepsy.

Pre-clinical studies involving rat [67] and canine [68] models of seizures demon-

strated that initiating vagus nerve stimulation during a seizure could terminate it or
lessen its severity. Clinical studies designed to quantify the therapeutic impact of

on-demand mode vagus nerve stimulation have also been conducted. Hammond [22]

recorded an EEG tracing illustrating the abrupt termination of an electrographic and

behavioral seizure following the initiation of on-demand VNS. Morris [25] retrospec-
tively analyzed seizure diaries from the E04 trial [32] and noted that 53% of patients

capable of receiving on-demand stimulation reported experiencing seizure termina-

tion or diminution. Similarly, Boon [4] noted that two-thirds of patients receiving
on-demand stimulation reported being able to interrupt seizures. While the results in
[25] and [4] are encouraging, it is important to note that they were determined using
patient and caregiver accounts of seizures as opposed to EEG tracings illustrating
seizure termination or diminution.

A significant proportion of VNS therapy patients depend on others to initiate

on-demand stimulation because the physical or cognitive symptoms of a seizure leave

them unable to do so themselves[4]. Depending on a caregiver to initiate on-demand

stimulation has two consequences: 1) it denies patients the therapeutic benefit of

on-demand stimulation in the absence of caregivers, and 2) it leads to an inconsistent

ability to terminate or attenuate seizures since caregivers may not always be able

114
to initiate on-demand stimulation immediately upon the clinical (let alone the elec-
trographic) onset of a seizure. Since there is anecdotal evidence suggesting that the
likelihood of affecting seizure progression decreases as the time between seizure onset
and the start of stimulation increases [22], this delay in detection may have significant
consequences.
A system that automatically initiates on-demand mode VNS following comput-
erized seizure onset detection could relieve caregivers, increase a patient's sense of
independence and security, and more frequently terminate or diminish seizure symp-
toms in individuals with seizures that respond to on-demand stimulation. The system
may also benefit patients who never realized they were on-demand responders since
neither they nor their caregivers were able to initiate stimulation following the onset
of a seizure.
In this chapter, we describe the design and clinical evaluation of a non-invasive
computerized system that automatically initiates on-demand VNS following the de-
tection of a prespecified seizure. The computerized system detects these events
through the patient-specific seizure detection methodology developed in Chapters
3 and 6.

7.2 Methods

7.2.1 System Overview

Figure 7-1 shows a block diagram of the computerized system; all the components
of the computerized system are external to the patient. The computerized system
is composed of a commercial acquisition system (Digitrace 1800 Plus from SleepMed
Inc.) that collects the EEG and ECG of a patient, a computer that analyzes the
EEG and ECG in real-time using the algorithms presented in Chapters 3 and 6, and
an electromagnet that is worn by the patient and positioned so that it rests over the
implanted VNS pulse generator.
When the computerized system detects the onset of a seizure, it energizes the

115
 ---------------------------------------------
Computerized System

VNS EEG and ECG EEG/ECG Based

Patient Acquisition Seizure Detection

Vagus Nerve Electromagnet

Stimulator

Figure 7-1: Block diagram of closed-loop vagus nerve stimulation system.

electromagnet worn by the patient. The magnetic field produced by the electromagnet
triggers the implanted generator to initiate on-demand stimulation of the vagus nerve.
The electromagnet initiates on-demand stimulation through the same mechanism that
is triggered when a permanent magnet is held over the implanted generator.

7.2.2 Study Protocol

The clinical evaluation of the computerized system followed a protocol approved

by the Committee on Clinical Investigations at the Beth Israel Deaconess Medical
Center (BIDMC), Boston, Massachusetts, USA, and the Massachusetts Institute of
Technology (MIT), Cambridge, Massachusetts, USA. Study participants were adult,
long-term users of VNS who reported no adverse effects from on-demand VNS. Par-
ticipants were admitted to the BIDMC General Clinical Research Center (GCRC) for
a period lasting up to five days. Admission to the GCRC took place after study staff
obtained informed consent, and confirmed that energizing the computerized system's
electromagnet reliably initiates on-demand stimulation. No changes to antiepileptic

drug (AED) regimens or VNS stimulation parameters were made during the study

period.

During the admission period Video, EEG, and ECG signals were recorded. After

116
the first 24 hours of the study, an electroencephalographer reviewed the collected
EEG and highlighted seizure events. Next, the computerized system was trained to
detect the identified seizure events using the EEG gathered in the first 24 hours of

the study. On subsequent days of the study, the computerized system was set to
automatically activate the participant's VNS generator whenever real-time detection
of the target seizure occurred.

7.3 Case Studies

In this section we examine three cases that involved the initiation of on-demand vagus
nerve stimulation in response to computerized seizure onset detection.

7.3.1 Patient A

Patient A Medical History

Patient A and Case 1 of Chapter 6 correspond to the same individual. Patient A is

a 39-year-old woman with a long history of refractory complex partial seizures. At
the time of admission to our study, she was experiencing 30-40 seizures per month.
Patient A's seizures last for 1-2 minutes and consist of repeatedly asking questions
and blank stares; her seizures are not accompanied by automatisms or tonic-clonic
movements. Following a seizure, patient A is confused for a few minutes and requires
reorientation by friends or family. Since Patient A does not experience a warning at
the onset of a seizure, she does not use the on-demand mode VNS.

Patient A Seizures

Figure 7-2 illustrates an EEG tracing of a typical seizure recorded from patient A.
The onset of the seizure, at 1486 seconds, involves rapid eye-blinking (eye flutter)

which manifests in the EEG as high-amplitude deflections on the channels {FP1 -

F7, FP1 - F3, FP2 - F8, FP2 - F7}. Coincident with the onset of eye-flutter is

an acceleration of the patient's heart rate, as seen on the bottom channel in Figure

117
7-2, and plotted in Figure 7-3. Finally, at 1492 seconds, 3-4 Hz theta waves appear
in the EEG on the temporal channel T4 - T6. Electrographic signs of the seizure
end around 1510 seconds, but clinical symptoms of the seizure persist for another 2-3
minutes.
Eye Flutter

Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4 I
F4-C4
C4-P4
P4-02
Fpl-F
F7-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

ECG
14 18 149 1492 1494 1496 1498 1500 1502 14 150 1508 151
Time (Seconds)

HR Acceleration Theta Waves

Figure 7-2: Example of a seizure within the scalp EEG of Patient A. The seizure,
which begins at 1486 seconds, involves rapid eye-blinking that results in downward
deflections on frontal EEG channels (e.g. {FP1 - F3, FP2 - F4}). Coincident with
the onset of rapid eye-blinking, the patient's heart rate accelerates as shown in Figure
7-3. Later, at 1492 seconds, a 3-4 Hz theta wave appears on the EEG channel T4-T6.

A total of 5 prerecorded seizures as well as two days of prerecorded awake and

sleep EEG were available to train a patient-specific detector to differentiate between
the seizure and non-seizure EEG of Patient A. The training seizure and non-seizure
data were recorded six months prior to patient A's admission to the study.

Patient-Specific, EEG-ECG-based Initiation of VNS

The earliest events that define Patient A's seizures are eye-flutter within the EEG and
a heart rate acceleration within the ECG. In order to reliably initiate VNS following
the coincidence of these events we used the detector architecture shown in Figure 7-4.
This architecture is similar to that developed in Chapter 6 except that the feature

118
 140 r

130"

120

110

S100- Seizure Onset

90 -

80

70
1460 1465 1470 1475 1480 1485 1490 1495 1500
Time (Seconds)

Figure 7-3: Seizure onset, at 1486 seconds, is associated with an acceleration of the
patient's heart rate.

vector used by the detector involves only spectral and spatial information extracted
from the EEG (no time evolution) and heart rate and heart rate change information
extracted from the ECG.

Figure 7-5 illustrates a seizure detected using the architecture illustrated in Figure
7-4. Seizure onset begins with rapid eye blinking at 992 seconds. The detector
declared seizure onset at 996 seconds, which is 4 seconds after the onset of eye-flutter
and 2 seconds before the onset of theta wave activity. Beginning at 1002 seconds a
spike-train appears on the "VNS" channel confirming the initiation of vagus nerve
stimulation. In this case, VNS was initiated 1 second after the onset of theta wave
activity. The heart rate acceleration that triggered the detector is shown in Figure

7-6; note its similarity to the training heart rate profile illustrated in Figure 7-3. The
computerized system automatically detected and initiated VNS on four other seizures

as described in [57].

119
 X1, 1
X ,1
2
EEG Ch 1 - i
X 8 ,1 * Test Feature Vector
* Training Seizure Feature Vectors
X , A Training Non-Selzure Feature Vectors
1 2
X
EEG Ch 2 2 ,2 A AA

X 8 ,2
X1, 1
6
EEG Ch 16 X2, 1
6

................ ... .............. Classification:

X 8 ,1 6
...................................... 2-Class Support Vector Machine
Mean Heart Rate X1,ECG
ECG e-
Heart Rate
Change X2,ECG
......................................

Feature Extraction

Figure 7-4: Block diagram of patient-specific seizure onset detector that combines
features extracted from the EEG and ECG signals.
Seizure Detected

Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02
VNS
ECG
988 990 992 994 996 998 1000 1002 1004 1006 1008 1010 1012 1014
Time (Seconds)

VNS Active

Figure 7-5: Initiation of VNS following computerized detection of the onset of a

seizure from Patient A. Seizure onset begins with rapid eye blinking at 992 seconds.
The detector declared seizure onset at 996 seconds, and initiated VNS in response.
Beginning at 1002 seconds, a spike-train appears on the "VNS" channel confirming
the initiation of vagus nerve stimulation.

120
 140

130

120

" 110-
Seizure Onset
S100
0a

90-

70

960 965 970 975 980 985 990 995 1000 1005 1010
Time (Seconds)

Figure 7-6: Seizure onset, at 992 seconds, is associated with an acceleration of the
patient's heart rate.

Comparing Stimulated and Non-stimulated Seizures

The computerized system did not initiate on-demand VNS following the onset of one

of patient A's seizures. This was done intentionally so that we could compare the
electrographic and behavioral character associated with stimulated seizures and a
non-stimulated seizure.

We did not detect a difference in the electrographic or behavioral character of

stimulated and non-stimulated seizures. Three possible explanations for this include:

1) the on-demand stimulus was not initiated early enough in the course of the seizure,
2) the on-demand VNS stimulus parameters (pulse current, frequency, and duration)
were not set appropriately 3) Patient A is not a responder to on-demand mode VNS.

7.3.2 Patient B

Patient B Medical History

Patient B is a 41-year-old woman with a long history of simple partial, complex partial

and secondarily generalized seizures. At the time of her admission to our study, she
was experiencing 5-6 seizures per month. Patient B's secondarily generalized seizures

121
consist of a tonic phase followed by a clonic phase, and finally a post-ictal phase. The
tonic-clonic phase lasts for 1 minute. The post-ictal phase, which is characterized by

generalized EEG slowing, lasts for 30 minutes. According to patient B's caregiver,
initiating VNS early in the ictal phase results in reduced anxiety and confusion during
the post-ictal phase.

Patient B Seizures

During the study period, patient B experienced two secondarily generalized seizures.

The first was used to train the computerized system, and the second to investigate
the impact of initiating on-demand VNS early in the course of patient B's seizure.

Figure 7-7 illustrates an EEG trace of the first seizure. The seizure begins following
the 1340 seconds with an electrodecrement most prominent on the occipital channels
{P3 - 01, P4 - 02}. Next, at 1348 seconds, a 3-5 Hz occipital rhythm emerges

from the electrodecrement and rapidly generalizes. By 1360 seconds, muscle activity

associated with the tonic-clonic phase of the seizure becomes visible. The tonic-clonic
phase of the first seizure lasted for 55 seconds and the post-ictal phase lasted for 29
minutes. During the post-ictal phase, the patient was unaware of her surroundings

and very anxious. The computerized system, using a detector of the form illustrated
in Figure 7-8, was trained to recognize the occipital rhythm and muscle activity
associated with the tonic-clonic phase of the seizure.

Patient-Specific, EEG-based Initiation of VNS

Figure 7-9 shows an EEG trace of the seizure that triggered the system to initiate
on-demand VNS. The seizure begins with an electrodecrement at 1485 seconds. The
computerized system failed to detect the onset of the focal occipital rhythm at 1490

seconds. Instead, at 1501 seconds, it detected the muscle activity associated with

the tonic-clonic phase of the seizure. Evidence of initiating on-demand VNS cannot

be seen near the time of seizure detection because the VNS channel is obscured by

muscle artifact. However, at the conclusion of the tonic-clonic phase, evidence of VNS

generator activity can be seen on the VNS channel as shown in Figure 7-10.

122
 Fp1-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fp1-FT

T3-T5
T6-Ol
Fp2-F8
F8-T4
T4-T6
T6-02

1336 1338 1340 1342 1344 1346 1348 1350 1362 1354 1356 1358 1360 1362 1364 1366 1368

Figure 7-7: Typical EEG change associated with onset of Patient B's seizure. The
seizure, which begins at 1340 seconds, involves an electrodecrement most prominent
on the occipital channels P3 - 01 and P4 - 02. Next, at the 1348 seconds, a 3-5
Hz occipital rhythm emerges from the electrodecrement and rapidly generalizes. By
1360 seconds, muscle activity associated with the tonic-clonic phase of the seizure
becomes visible.

. .................................... X 1,1

E E G Ch 1 ...............................................................
X8,1 Test Feature Vector
..... 0 Training Seizure Feature Vectors
X1, 2 A Training Non-Seizure Feature Vectors

1,16A
X8,2 A
AA A,&

EEG Ch 16 --- / ' l AAAA

A A "-
- 2,16 AA

SX8,16 Classification:
2-Class Support Vector Machine

Figure 7-8: Block diagram of patient-specific seizure onset detector that uses features
extracted from the EEG signal.

123
 Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

VNS

1479 1481 1483 1485 1487 1489 1491 1493 1495 1497 1499 1501 1503 1505 1507 1509

Figure 7-9: Initiation of VNS following computerized detection of the onset of a

seizure from Patient B. The seizure begins with an electrodecrement at 1485 seconds.
At 1501 seconds, the computerized system detected the muscle activity associated
with the tonic-clonic phase of the seizure and initiated VNS in response. Evidence
of VNS generator activity can be seen on the VNS channel at the conclusion of the
seizure as shown in Figure 7-10.

Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4 ~
~C-~-c~
I-
F4-C4 - CL~_~_I~C I~L ---- h ~-- _ --
C4-P4 1IC------
~--l--------rY- -
P4-02
FpI-F7
F7-T3
T3-TG
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

VNS

Figure 7-10: The spike-train seen on the VNS channel between 1555-1565 seconds
confirms the automatic initiation of vagus nerve stimulation following the onset of
the seizure illustrated in Figure 7-9.

124
Comparing Stimulated and Non-stimulated Seizures

Similar to the first seizure, the tonic-clonic phase of the second seizure lasted for 65
seconds and the post-ictal phase for approximatly 30 minutes. We believe that the

length and character of these phases was unaltered because the computerized system
initiated on-demand VNS during the generalized (tonic-clonic) portion of the seizure
rather than at the focal stage of the seizure. However, behaviorally, relative to the
first seizure, patient B was significantly less anxious during the post-ictal phase. Our
observation that applying on-demand VNS during the seizure improves post-ictal

recovery from the seizure is consistent with the experience of patient B's caregiver.

7.3.3 Patient C

Patient C Medical History

Patient C and Case 2 of Chapter 6 correspond to the same individual. Patient C is

a 24 year-old man with Lennox-Gastaut syndrome. At the time of his admission to
our study, he was experiencing daily, generalized tonic seizures. The tonic phase of
the seizure lasts between 10-20 seconds and the post-ictal phase for approximately an
hour. Since Patient C does not experience a warning prior to the onset of a seizure
he does not use the on-demand mode of VNS.

Patient C Seizures

Figure 7-11 illustrates an EEG trace of a typical seizure recorded from Patient C.
The onset of the seizure, at 248 seconds, involves an electrodecrement that includes
all EEG channels and lasts for 14 seconds. During the electrodecrement the patient
experiences generalized, tonic muscle contractions. At 262 seconds, both the elec-
trodecrement and tonic contractions end and the EEG exhibits generalized rhythmic

activity. The computerized system, using a detector of the form illustrated in Figure

7-8, was trained to recognize the electrodecrement associated with the tonic phase of
the seizure.

125
 Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

Figure 7-11: Example of a seizure within the scalp EEG of Patient C. The seizure,
which begins at 248 seconds, involves a 14 second period of low-amplitude EEG
activity across most EEG channels. Later, at 262 seconds, generalized, rhythmic
activity develops.

EEG-based Initiation of VNS

Figure 7-12 shows a seizure that triggered the system to initiate on-demand VNS.

The system noted seizure activity at 59 seconds (3 second latency) and initiated
vagus nerve stimulation in response. Beginning at 69 seconds, one can see a spike-
train on the VNS channel indicating successful initiation of vagus nerve stimulation;
the spike-train can be seen to continue in Figure 7-13.

Comparing Stimulated and Unstimulated Seizures

We did not detect a change in the electrographic or behavioral character of Patient

C's seizure following the initiation of on-demand VNS. Three possible explanations
for this include: 1) the on-demand stimulus was not initiated early enough in the

course of the seizure 2) the on-demand VNS stimulus parameters (pulse current,

frequency, and duration) were not set appropriately 3) Patient C is not a responder

to on-demand mode VNS.

126
 Fpl-F3
F3-C3
C3-P3
P3-01
Fp2-F4
F4-C4
C4-P4
P4-02
Fpl-F7
F7-T3
T3-T5
T5-01
Fp2-F8
F8-T4
T4-T6
T6-02

VNS
4 51 53 55 57 59 r VW !q69
Time (Seconds)

Figure 7-12: Initiation of VNS following computerized detection of the onset of a

seizure from Patient C. The seizure begins with an electrodecrement at 56 seconds.
The computerized system noted seizure activity at 59 seconds and initiated vagus
nerve stimulation in response. Evidence of VNS generator activity can be seen on the
VNS channel starting at 69 seconds.

Figure 7-13: The spike-train seen on the VNS channel (first channel from the bottom),
between 68-84 seconds, confirms the automatic initiation of vagus nerve stimulation
following the onset of the seizure illustrated in Figure 7-12.

127
128
Chapter 8

Seizure-Triggered Single Photon

Emission Computed Tomography

In this chapter we illustrate the feasibility of applying the seizure detection methodol-
ogy developed in Chapter 3 to the initiation of a delay-sensitive diagnostic application.
More specifically, the infusion of a neuroimaging radiotracer following seizure onset.

8.1 SPECT in Epilepsy

Surgical management of medically intractable seizures requires the identification and

resection of an individual's seizure focus, the cerebral site that gives rise to a seizure.
Defining the seizure focus usually involves electrical characterization of seizures through
EEG, anatomical imaging of the brain through magnetic resonance imaging (MRI),
and functional imaging of the brain during the seizure and non-seizure states using
Single Photon Emission Computed Tomography (SPECT) [5].
A SPECT image reflects differences in blood flow, or perfusion, to various regions
within the brain. Since the seizure focus tends to be hypermetabolic during a seizure,
demanding a greater share of cerebral blood flow relative to surrounding regions, a
SPECT image taken at the time of a seizure (an ictal SPECT) reveals the seizure
focus as a region of hyperperfusion as shown in Figure 8-1. In order to obtain an ictal
SPECT image that highlights the seizure focus well, a radiotracer must be injected

129
 Figure 8-1: SPECT Image taken outside the seizure state (Interictal) and during
the seizure state (Ictal I). During a seizure the seizure focus is hyperperfused and
appears as a bright spot in a SPECT Image. The third image (Ictal II) is the result
of subtracting the first two images.

into a patient soon after the electrical onset of a seizure [63]. To do this one could
have: 1) an EEG technician continuously monitor a patient's EEG in real-time in

order to rapidly detect and declare the electrical onset of a seizure 2) a trained nurse
remain near the patient so that the correct radiotracer dose is safely and promptly

administered following declaration of seizure onset by the EEG technician.

Such an approach is costly and nearly impossible to do well when multiple patients
need an ictal SPECT within a hospital's Epilepsy Monitoring Unit. In practice, a

trained nurse uses a syringe to inject the radiotracer dose after a caregiver near the

patient observes the clinical manifestations of a seizure. This results in appreciable

injection delays because the seizure's clinical onset tends to lag its electrographic

onset; early signs of the seizure's clinical onset are subtle; and the trained nurse is far
away from the patient. In our experience injections are started 10-130 seconds after
seizure onset as shown in Figure 8-2. Large delays lead to poor localization of the
seizure focus due to the visualization of secondarily activated foci in addition to the
primary seizure focus.

A system capable of automatically detecting the electrical onset of a seizure and

dispensing the appropriate dose of radiotracer using a drug infusion pump could result

in shorter and less varied radiotracer injection delays. Overcoming the technical and

logistical challenges inherent to the practice of ictal SPECT may increase both the

utility and efficacy of ictal SPECT as a tool for seizure focus localization. This will

130
 Ictal SPECT Radiotracer Injection Delay
From Electrographic Seizure Onset

140

130

120
o 110
S100

70

60
50
4 =

30

10

Figure 8-2: Delay associated with injecting the ictal SPECT radiotracer ranges be-
tween 10-130 seconds for 18 pediatric patients undergoing ictal SPECTs at Children's
Hospital Boston.

particularly benefit epilepsy surgery candidates dependent on functional neuroimag-

ing for the definition of seizure foci that are difficult to visualize using anatomical

imaging modalities such as MRI.

In this chapter we describe the design and clinical evaluation of a computerized

system that automatically initiates radiotracer infusion following the detection of the
electrical onset of a seizure. The computerized system detects seizure onset through
the patient-specific methodology developed in Chapter 3. Other automated ictal
SPECT systems [15, 53] did not involve computerized detection of the electrical onset
of a seizure. Instead, the clinical onset of a seizure prompted a health care provider to

remotely initiate radiotracer infusion using a computer-controlled drug pump. When

the system described in [15] was evaluated on 26 patients, radiotracer injection delays

ranged between 3-48 seconds.

 8.2 Methods

8.2.1 System Overview

Figure 8-3 shows a block diagram of the computerized system. The computerized

system is composed of a commercial acquisition system (BioLogic Inc.) that collects

the EEG of a patient, a computer that analyzes the EEG in real-time using the

patient-specific algorithm developed in Chapter 3, and a computer-controlled drug

pump that dispenses the ictal SPECT radiotracer.

------------------------------------------
Computerized System

Ictal SPECT EEG

Patient EEG Acquisition Seizure Detection

Drug Infusion
Pump

:------------------------------------------

Figure 8-3: Block diagram of system for automatic infusion of ictal SPECT radio-
tracer.

When the computer detects the onset of a seizure through the analysis of the EEG
signal stream, it computes the appropriate dose of radiotracer to infuse. The dose is
a function of how long after radiotracer preparation the seizure occurs. Finally, the
computer issues a command to the drug pump to dispense the radiotracer dose as a
bolus into the patient.

8.2.2 Study Procotol

The clinical evaluation of the computerized system followed a protocol approved by

the Institutional Review Board at Children's Hospital Boston (CHB), Boston, Mas-

sachusetts, USA. The protocol involved two groups that were simultaneously active,

132
but which were blinded to each other's activities. The first group (Group 1) included
a patient scheduled for an Ictal SPECT as part of his/her hospital admission, and
the medical staff responsible for carrying out that ictal SPECT according to CHB
protocol. The second group (Group 2) included the computerized system. The two
groups are shown in Figure 8-4.

GROUP #1: Routine Ictal SPECT Study Medical Staff, Patient I

I
EEG
Ictal SPECT Study
I Patient
I

EEG
Medical Staff waits to inject radiotracer upon
notification of seizure onset by caregiver.

GROUP #2: Computerized System

Seizure Detector

I
F4 Water (radiotracer surrogate)

Plastic Receptacle (patient surrogate) I

Figure 8-4: Protocol for evaluating automatic system for infusion of ictal SPECT
radiotracer. See section 8.2.2 for details.

Group 1 carried out a routine ictal SPECT procedure in accordance with the CHB
ictal SPECT protocol. That protocol involved inserting an intravenous line into the
patient for radiotracer delivery; recording live scalp EEG from the patient as shown
in link 1 of Figure 8-4; awaiting notification of the clinical onset of the patient's
seizure by an attending caregiver; confirming seizure activity on the recorded EEG;

133
and finally mobilizing a trained nurse to the patient's room to deliver the appropriate
radiotracer dose as well as a volume of saline to flush the intravenous line connected
to the patient.
The scalp EEG captured from the patient in Group 1 was forwarded in real-time
to the computerized system in Group 2 as shown by links 1 and 2 of Figure 8-4. The
system processed the EEG using the patient-specific detector discussed in Chapter
3. The detector was trained using seizure and non-seizure data recorded between the
time of admission and the ictal SPECT study. Upon detecting the electrographic
onset of a seizure, the system determined the appropriate dose to be delivered and
instructed the drug pump to initiate radiotracer delivery as shown in links 3 and 4 of
Figure 8-4. The pump delivered a volume of water (a surrogate for the radiotracer)
into a plastic receptacle (a surrogate for the patient). The volume of water equaled
the sum of the appropriate radiotracer dose and the volume of saline injected by the
nurse to flush the patient's intravenous line.

8.3 Results of Clinical Evaluation

We evaluated the performance of the computerized system during eight ictal SPECT
trials. Figure 8-5 shows the time elapsed between electrographic seizure onset and
dispensing of the radiotracer dose for both the computerized system (column 2) and
the ictal SPECT clinical team (column 3); also shown in Figure 8-5 is the delay with
which the computerized system first recognizes the electrographic onset of a seizure
(column 1).
We find that our computerized system recognized the electrographic onset of a
seizure within 5.1 ± 2.8 seconds and that it completed dispensing the appropriate
radiotracer volume within 19.3 ± 2.3 seconds. We attribute the large delay between
recognition of seizure onset and pump infusion of the requisite volumes to the slow
rate with which the pump selected for the study dispenses fluid. A faster pump would
have resulted in a shorter delay. The ictal SPECT clinical team completed infusion of
the radiotracer within 27.7 ± 8.5 seconds. For patient number 7, the first seizure went

134
unnoticed by the clinical team (absent column 3 of Figure 8-5), but was detected by
the computerized system. Missing the first seizure of patient 7 prompted the need
for a second ictal SPECT study, which both the clinical team and system detected.

Automatic vs Manual Ictal SPECT

Radiotracer
U PumpCompletes Injection andFlush 0 NurseCompletesRadiotracerInjectionand Flush

1 2 3 4 5 6
Patient Number

Figure 8-5: Comparison of delay between seizure onset and completion of radiotracer
infusion by the clinical team (Group 1) and the computerized system (Group 2).

The computerized system incorrectly dispensed the surrogate radiotracer on 4

occasions over the course of 33 hours. Figure 8-6 shows the latency with which each
of the eight seizures were detected (blue bars) as well as the number of false positives
(FP) and total number of hours processed per ictal SPECT study (text above blue
bars). Two of the eight ictal SPECT trials accounted for all the false detections
generated by the computerized system. This result suggests that with careful patient
selection the system can be used to initiate radiotracer infusion without concern that
injections stem from false detections.

135
 Automatic Seizure Detector Performance

20

18

16

=14
o
212
a.

"8
E
20
U.

2 3 4 5 6
Patient Number

Figure 8-6: Latency and number of false detections declared by the computerized
system during eight Ictal SPECT studies.

136
Chapter 9

Patient-Specific Seizure Onset

Detection using iEEG

In this chapter we adapt the patient-specific detector developed in Chapter 3 for the
purpose of detecting the onset of a seizure using the intracranial Electroencephalo-
gram (iEEG). Since the preferred embodiment of an intracranial seizure onset detector
involves implementation on a low-power, computationally-constrained device, we for-
mulate the algorithm with this constraint in mind. In this chapter we also compare
the performance of our algorithm to that of a state-of-the-art, patient non-specific
intracranial seizure onset detector.

9.1 Why Detect Seizure Onset Using iEEG?

As discussed in Chapter 2, the intracranial Electroencephalogram provides a spatial

and temporal summary of the electrical activity of a population of neurons. However,
unlike the scalp EEG, intracranial EEG provides a summary with higher spatial
resolution since each iEEG electrode samples the activity of a smaller population
of neurons. The higher spatial resolution of iEEG allows one to notice the neuronal

hypersynchrony associated with a seizure tens of seconds before the same phenomenon
is noticeable within the scalp EEG [39]. At the same time, the higher spatial resolution

of iEEG permits the recording of a wider gamut of abnormal, non-seizure activity that

137
is not visible within the scalp EEG [26, 59, 62].
An implantable medical device capable of detecting and reacting to the onset of a
seizure within the iEEG can facilitate applications that are not possible using devices
lacking seizure detection capabilities. The device could 1) alert patients of the electri-
cal onset of a seizure before the development of clinical symptoms 2) align a stimulus
with the onset of a seizure to suppress the seizure 3) maintain an account of seizure
activity so that physicians can objectively determine the efficacy of a therapeutic
device.
Performing rapid and reliable seizure onset detection within an implantable med-
ical device is challenging. Algorithms for detecting seizure onset within iEEG signals
must be capable of detecting seizures given substantial variability of seizure charac-
teristics across patients, and overlap between the features of seizure and non-seizure
activity within the iEEG of a patient. At the same time, such algorithms need to
consume a small amount (- 50W) of the total power budget of the medical device.

9.2 Patient-Specific Detector Architecture

The implantable neurostimulator on which we sought to implement a seizure onset

detector is described fully in [2]. This neurostimulator is composed of an analog front-
end and a digital back-end. The analog front-end is capable of processing two iEEG
channels. For each channel, the analog front-end can extract the spectral power within
two bands with configurable center frequency and bandwidth. The digital back-end,
which has limited computational capabilities, can be used to implement decision logic
that ascertains the presence or absence of seizure activity.
The architecture of our patient-specific detector for the hardware in [2] is illus-
trated in Figure 9-1. The detector configures the neurostimulator's analog front-end
to extract, from each of two iEEG channels, the power within the frequency bands
0-16 Hz and 15-37 Hz. Next, within the neurostimulator's digital back-end, the detec-
tor samples the spectral power produced by the front-end and forms a 4-dimensional
feature vector that encodes spectral and spatial dependencies between the input chan-

138
nels. Finally, the detector classifies the observed feature vector using a support-vector
machine (SVM) trained to differentiate between a patient's seizure and non-seizure
activity using parameters y = 0.1, J = 17, and a value of C left at the default value

of the SVMLight software package [28]. Implementation of a support-vector machine

with a nonlinear decision boundary required simplification of the SVM discriminant

function using the methodology discussed in the following section.

IEEG
Channel 1 r1 -1

I,"
I
0-a, r *
x1, %o
OD
12

0,

IEEG
Channel2 X 6.2

0 02 0.4 06 0. 1
rwgy0.I11lIz

Feature Extraction Feature Vector Classification Stage

Figure 9-1: Block diagram of a patient-specific seiz ure detection algorithm that uses
spectral and spatial features extracted from two iE EG channels.

9.2.1 Cost of Support-Vector Machine Classification

When the support-vector machine learning algorithm is used to determine a linear

decision boundary, the resulting discriminant function (equation 9.1) requires little
computation to implement. In the case of the detection architecture shown in Figure
9-1, this discriminant function would involve 4 addition and 4 multiplication opera-
tions.

f(X) = Seizure if WTX + f > O (9.1)

f(X) = Non-seizure if WTX + 3 < 0

In the case of a nonlinear boundary, determined using a radial basis kernel, the
resulting discriminant function (equation 9.2) is more complex. The discriminant

139
function contains a number of terms, Nsv, equal to the number of support-vectors de-
termined by the support-vector machine learning algorithm. The number of support-
vectors is partly governed by the complexity of the classification task. As the similar-
ity between an individual's seizure and non-seizure activity increases, more support-
vectors are needed in order to define a more complex decision boundary, and as a
result the computational cost of equation 9.2 increases. In our experience, for a de-
tector with the architecture illustrated in Figure 9-1, Nsv ranges between 50 and
300. This precludes implementation of the discriminant function in equation 9.2 on
the hardware described in [2].

Nsv

f(X) = Seizure if {- a, exp(- lX - X,112)} + 0 > 0 (9.2)

Z=1
Nsv
f(X) = Non-Seizure if {- a, exp(-7|lX - X,112 )} + 0 < 0

In the following section, we broadly outline the steps involved in a model order
reduction technique developed by Scholkopf [52]. This technique allows the nonlinear
discriminant function in equation 9.2 to be expressed using Msv < Nsv terms.

Reducing The Number of Support-Vectors

Conceptually, the support-vector machine determines a nonlinear boundary in the

space of feature vectors by determining a maximum margin, linear boundary in the
higher-dimensional space induced by the kernel. Denote a feature vector in the feature
space as X, and its image in the higher-dimensional space as O(X,) where 0(.) is a map
from the feature space to the higher-dimensional space. In the higher-dimensional
space the linear boundary is defined by its normal vector W, which can be expressed
as a linear combination of the higher-dimensional image of Nsv support-vectors as
shown in equation 9.3.

140
 Nsv

W = ao(X) (9.3)

In order to obtain a nonlinear decision boundary in the feature space that uses

Msv < Nsv terms, we seek a normal vector W that is close to W in the sense that

1 W - W 112is small. Moreover, W should be expressed as a combination of the

higher-dimensional image of Ms new support-vectors S, as shown in equation 9.4.
Since the L 2-norm of the difference between W and W is being minimized, the higher

dimensional images of feature vectors need not be computed because expansion of the
expression being minimized yields inner-products of higher-dimensional vectors that
can be evaluated using a kernel.

Msv

-=77,S (9.4)
2=1

The Msv new support vectors S, are obtained through a sequential process. The
first, scaled vector r~S 1 is chosen as the best single vector approximation of W through
solving the following minimization problem

1 = Mim
71iS q(S) - W 112 (9.5)

The next scaled vector / 2S 2 is chosen as the best single vector approximation of
the residual vector Y = W - q S 1 through solving the following optimization problem

min 11q(S) - Y 112

= m2S2 (9.6)

This process is repeated until the set of new support vectors 71S, i = 1... Msv

is complete. With the new set of support vectors we can reexpress the discriminant

141
function in equation 9.2 using Msv instead of Nsv terms as shown in equation 9.7.

Msv
f(X) = Seizure if { I,2exp(-y|X - Sj 2
)} + 0 > 0 (9.7)
Z=l
MSV
f(X) = Non-Seizure if {E 1 exp(--ylX - S'l12)} + 0 < 0

As an example, Figure 9-2 shows the superposition of a nonlinear boundary re-

quiring Nsv = 50 terms (solid line) and an approximation of that boundary using
Msv = 8 terms (dashed line). In our evaluation of the patient-specific detector
described in section 9.2 we restricted equation 9.7 to M=8 terms.

O o0 . "
S0.6 0

0.2 -

0 0.2 0.4 0.6 0.8 1

Energy (0-16Hz)

Figure 9-2: Superposition of a nonlinear SVM boundary requiring Nsv 50 terms

(solid line) and an approximation of that boundary using Msv = 8 terms (dashed
line).

9.3 Patient Non-specific Detector Architecture

The patient non-specific detector evaluated in this chapter is a simplification of the

Osorio-Frei seizure detector [38]. The architecture of this detector is shown in Figure
9-3. The detector monitors, for each channel, the ratio of current and background
energy in the 15-37 Hz frequency band; the current energy is computed using a window
of 2 seconds while the background energy is computed using a window of 30 minutes.

142
 Whenever the ratio of foreground to background energy on any channel exceeds a
threshold R for T seconds a seizure is declared.

Foreground
IEEG " Energy Ratio>R
Channel 1 For
"---. [ Background TSeconds
Energy

Foreground
IEEG Energy I Ratio R
Channel,2 For
Background TSeonds
... Energy

Figure 9-3: Block diagram of patient non-specific seizure detection algorithm.

9.4 iEEG Data Set

The data set used to evaluate the performance of the two detectors included 81
hours of intracranial EEG collected from 17 adult subjects. On average, 4.5 hours
of recording time containing 3 seizures were available per patient. For each patient,
an electroencephalographer identified the onset time of all seizures as well as the
two iEEG channels demonstrating the earliest signs of seizure activity. These two
channels were the only channels processed by both algorithms.

9.5 Performance Comparison

The performance of the patient-specific detector using both linear and nonlinear
SVMs was compared to the performance of the patient non-specific detector set to
declare a seizure following T = 3 seconds and T = 10 seconds.

Figure 9-4 illustrates the average latency with which the patient-specific and non-

specific detectors declare the onset of a seizure for each of the 17 subjects. The

143
patient-specific detector that used a nonlinear SVM (reduced to use MNSV=8 terms)
detected 60/61 seizures within 8.5 ± 5.1 seconds. The same detector using a linear
SVM detected 60/61 seizures within 9.3 ± 4.8 seconds. The patient non-specific
algorithm detected 55/61 seizures within 10.5 ± 8.7 seconds. Note that the patient
non-specific detector failed to detect the seizures of Patient 3, which accounts for the
absence of a bar.
O Patient-Specific Nonlinear SVM E Patient-Specific Linear SVM 1Patient Non-Specific T=3 Sec
40

35

30

S25

20
20

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Patient Number

Figure 9-4: Comparison of the detection latencies of the patient-specific and patient
non-specific seizure detectors. The patient-specific detector that used a nonlinear
SVM detected seizures within 8.5 ± 5.1 seconds. The same detector using a linear
SVM detected seizures within 9.3 + 4.8 seconds. The patient non-specific algorithm
(with T=3) detected seizures within 10.5 ± 8.7 seconds.

Figure 9-5 illustrates the number of false detections declared by the detectors for
each of the 17 test subjects (40 hours of test data). The patient-specific detector
employing a nonlinear SVM made 19 false detections; the same detector using a
linear SVM made 28 false detections. The patient non-specific detector made 126
false detections.
When the patient non-specific detector was set to declare a seizure following T=10
seconds, it made 17 false detections as illustrated in Figure 9-6 alongside the patient-
specific method's false detections. Furthermore, with T=10 seconds the patient non-
specific algorithm detected 41/60 seizures within 18.7 + 10.8 seconds as shown in

144
 O Patient-Specific Nonlinear SVM 0 Patient-Specific Linear SVM a Patient Non-Specific T=3 Sec
50

45

35 ----
------------

E
15
25
2
10

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Patient Number

Figure 9-5: Comparison of the number of false detections declared by the patient-
specific and patient non-specific seizure detectors during 40 hours of non-seizure data.
The patient-specific detector with a nonlinear SVM declared 19 false detections. The
same detector using a linear SVM declared 28 false detections. The patient non-
specific detector (with T=3) declared 126 false detections.

Figure 9-7 alongside the patient-specific detector latencies.

This comparison demonstrates that the patient-specific approach detects a larger

percentage of seizures with both a lower false-detection rate and smaller detection
latency when compared to the patient non-specific approach.

9.6 Reduced and Non-Reduced Support-Vector Ma-

chines

In this chapter the number of support-vectors used in equation 9.7 was reduced to
AMsv = 8 using the method developed by Scholkopf [52]. In this section we com-
pare the performance of discriminant functions that use different number of support-

vectors in order to show that little performance is lost as a result of reducing the

number of support-vectors used to classify an observed feature vector.

Figures 9-8 and 9-9 compare the performance of support-vector machines that

145
 O Patient-Specific Nonlinear SVM 0 Patient-Specific Linear SVM a Patient Non-Specific T=10 Sec
12

10

.2

Z 2
4 ....
.+
.........

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Patient Number

Figure 9-6: Comparison of the number of false detections declared by the patient-
specific and patient non-specific seizure detectors during 40 hours of non-seizure data.
The patient-specific detector with a nonlinear SVM declared 19 false detections. The
same detector using a linear SVM declared 28 false detections. The patient non-
specific detector (with T=10) declared 17 false detections.
O Patient-Specific Nonlinear SVM a Patient-Specific Linear SVM a Patient Non-Specific T=10 Sec
60

50 ------

40

20

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Patient Number

Figure 9-7: Comparison of the detection latencies of the patient-specific and patient
non-specific seizure detectors. The patient-specific detector that used a nonlinear
SVM detected seizures within 8.5 + 5.1 seconds. The same detector using a linear
SVM detected seizures within 9.3 ± 4.8 seconds. The patient non-specific algorithm
(with T=10) detected seizures within 18.7 + 10.8 seconds.

146
use Msv = 8 and Msv = 16 support-vectors with a support-vector machine that
is allowed to use all support-vectors. On average, the support-vector machine that
used all support-vectors required 321 support-vectors. Figure 9-8 illustrates that
significantly reducing the number of support-vectors does not impact a detector's
latency, and Figure 9-9 shows that the support-vector machines with fewer support-
vectors declared only a few extra false detections.

08 Support Vectors 16 Support Vectors nAIl Support Vectors

25

20

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Patient Number

Figure 9-8: Comparison of latency of detectors that use support-vector machines with
different number of support-vectors

08 Support Vectors i16 Support Vectors NAII Support Vectors

4:1

v
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Patient Number

Figure 9-9: Comparison of number of false detections declared by detectors that use
support-vector machines with different number of support-vectors

147
 9.7 Case Studies

The following examples illustrate why a patient-specific approach can yield better

seizure onset detection performance.

9.7.1 Latency

I I I I I I i I i I I I i I I I I I I I
1500
Seizure Onset Patinet-Specific Patinet
Non-Specific

1000

500 I L it h . -

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44
Time (Seconds)

Figure 9-10: Example of a seizure within the iEEG of Patient 6. The seizure, which
begins at 2 seconds, consists of a few spikes that evolve into a high-amplitude spike
train. The patient-specific detector recognized the seizure at 16 seconds, while the
patient non-specific detector did so at 40 seconds.

Figure 9-10 shows an iEEG tracing of a typical seizure from Patient 6. The
seizure, which begins at 2 seconds, consist of a few spikes that evolve into a high-
amplitude spike train. The patient-specific detector recognized the seizure at 16
seconds, while the patient non-specific detector did so at 40 seconds. Presumably,
the patient non-specific detector had a long latency because the dominant frequency
component present at seizure onset (1 Hz) is outside the frequency band monitored by

the patient non-specific detector (15-37 Hz). The patient-specific detector was able

to detect the seizure rapidly because it learned from training seizures the importance

of low frequency information.

148
9.7.2 False Detections

Figure 9-11 shows an iEEG tracing of a seizure from Patient 5. The seizure, which
begins at 6 seconds, consist of a high-frequency rhythm that increases in amplitude
and decreases in frequency as the seizure progresses. Both the patient-specific and

patient non-specific detectors could rapidly detect the onset of this patient's seizure.

2500 i I I

2000
1500N

1000 1

500

0 , 123

0 5 10 15 20 25 30 35 40 45
Time (Seconds)

Figure 9-11: Example of a seizure within the iEEG of Patient 5. The seizure, which
begins at 6 seconds, consists of a high-frequency rhythm that increases in amplitude
and decreases in frequency as the seizure progresses.

During the non-seizure state, the iEEG of Patient 5 contained bursts of rhythmic

activity such as that seen in Figure 9-12. The patient-specific detector did not declare
these bursts as false detections because it had learned, through training, that they are
a feature of the patient's baseline activity. On the other hand, the patient non-specific
detector declared many of these bursts as seizure events even though they have spatial
and spectral characteristics that differ significantly from those of the actual seizure.

9.8 Predicting Clinical Seizure Onset

Detecting the electrical onset of a seizure using iEEG could be used to alert a patient

to a seizure prior to the development of debilitating clinical symptoms. The success

of this application depends on the separation between the the electrical onset of a

seizure within iEEG and the onset of clinical symptoms. This separation varies across

149
 W &A4
*MAM4M

4821 4822 4823 48 4825

Time (Seconds)

Figure 9-12: Example of a burst of rhythmic non-seizure activity within the iEEG
of Patient 5. The patient non-specific detector declared this burst as a seizure event
even though its spatial and spectral character differs significantly from that of the
seizure shown in Figure 9-11.

patients. Figure 9-13 shows when, relative to the clinical onset, the patient-specific
and patient non-specific detectors declare the onset of a seizure for each of the 17 test
subjects. Negative latencies correspond to detections that preceded the clinical onset
of a seizure and positive latencies correspond to detections that followed the clinical
onset. An alarm raised by our patient-specific detector could provide patients 3 and
6 more than thirty seconds to prepare for the clinical onset of a seizure.

9.9 Implementation

Both the patient-specific and non-specific algorithms were implemented on the neu-
rostimulator described in [2]. As discussed in section 9.2, the analog front-end of
the stimulator is capable of processing two iEEG channels, and for each channel, ex-
tracting the spectral power within two bands with configurable center frequency and
bandwidth. The digital back-end samples the analog power profile, assembles these
samples into a feature vector, and classifies the feature vector using a support-vector
machine. This signal chain is illustrated the Figure 9-14.
By delegating the feature extraction to the analog front-end, the digital back-end
remains idle until it is necessary to sample the power profile and test for the presence

150
 u P atlent-Specific Nonlinear SVM U Patient Non-specific T=10 sec

304

20 . ... .
--.----------..

20

5 i 7 8 9 10 11 12
-10

-20 - A
_,i ..

Patient Number

Figure 9-13: Detection delay relative to the clinical onset of a seizure. An alarm
based on detecting the electrical onset of a seizure using our patient-specific method
could provide some patients, such as patients 3 and 6, enough time to prepare for the
clinical onset of a seizure.

Analog:
Senses IEEG and Extracts Spectral Power

IEEG Channel 1

.... ..............
%.. ..

........ .........

IEEG Channel 2

Digital:
Samples Spectral Power, Creates Feature Vector, and Classifies

X1, 1 _i
f (X) > 0 -x= 1 ,2
xL
X2,1

Lx2,2j

Figure 9-14: Implementation of our machine-learning based, patient-specific detector

on the hardware described in [2]. An analog front-end processes two iEEG channels,
and for each channel, extracts the spectral power within two configurable frequency
bands. A digital back-end samples the analog power profile, assembles these samples
into a feature vector, and classifies the feature vector using a support-vector machine.
of seizure activity. This sampling can occur at a rate as low as 1 Hz. Had feature
extraction been delegated to the digital back-end, then sampling would have had to
occur at a rate suitable for iEEG signals, which is typically 200-250 Hz.
Our measurements indicate that the total current consumed by the patient non-
specific detector was 32 1 A. The patient-specific detector consumed 12PA when using
a linear SVM and 56jiA when using a nonlinear SVM. These measurements suggest
that the patient-specific detector employing a linear SVM exhibits the most favorable
combination of accuracy, latency, and power consumption.

152
Chapter 10

Conclusion and Future Work

In this chapter, we conclude the thesis with a summary of its goals and contributions

followed by proposed improvements and directions for future work.

10.1 Goals and Contributions

The goal of this thesis was to design, evaluate, and clinically test a seizure onset
detection algorithm. The algorithm we developed contributes technically to the field
of seizure detection in the following ways:

" Enhanced Performance: In Chapter 3 we presented a machine-learning

based, patient-specific seizure onset detector. The detector uses a support-
vector machine to classify a feature vector that automatically encodes the time
evolution of spectral and spatial features within the scalp EEG. When trained
on 2 or more seizures and tested on 844 hours of continuous scalp EEG from
23 pediatric subjects, our algorithm detected 96% of 163 test seizures with a
median detection delay of 3 seconds (average 4.6 seconds) and a median false
detection rate of 0.07 false detections per hour (average 0.13 false detection per

hour).

* Uses Multiple Physiologic Signals: In Chapter 6 we showed how the de-

tector presented in Chapter 3 can be extended with information from another

153
 physiologic source whenever the scalp EEG alone is an unreliable indicator of
seizure onset. This capability is important for the detection of seizures whose
onsets lack the development of rhythmic activity and instead reflect physical
sequelae of the seizure.

* Minimal User Intervention: Our algorithm does not require a user to define
the values of parameters such as which EEG channels to monitor for seizure
activity. In our approach, a user only needs to define the onset of activity asso-
ciated with a seizure in a set of physiologic signals. The relationships between
these signals that distinguish the seizure and non-seizure periods are automat-
ically learned.

* Suitable for Implantable Medical Devices: In Chapter 9 we showed that

the algorithm presented in Chapter 3 can be adapted for the detection of seizure
onsets within intracranial EEG, and can be implemented on the low-power
hardware of an implantable neurostimulator. When evaluated on 81 hours of
intracranial EEG containing 61 seizures and gathered from 17 adult subjects,
our algorithm detected 60/61 seizures within 8.5 seconds and declared a total
of 19 false detections.

In this thesis we also used our algorithm to enable delay-sensitive therapeutic

and diagnostic applications:

* Non-invasive Closed-Loop Control of the Vagus Nerve Stimulator: In

Chapter 7 we showed how our algorithm can be used to initiate vagus nerve
stimulation in response to detecting the onset of a seizure using multiple phys-
iologic signals. As an example of the system's capabilities, during an 81 hour
clinical test of the system on a patient, the computerized system detected 5/5
seizures and initiated VNS within 5 seconds of the appearance of ictal discharges
in the EEG.

* Computerized Initiation of Ictal SPECT Studies: In Chapter 8 we

showed how our algorithm can be used to initiate a functional neuroimaging

154
 study following seizure onset. The neuroimaging modality, ictal SPECT, is
used to radiographically localize the cerebral origin of a seizure. Our system
could initiate injection of the radiotracer used for ictal SPECT within 19.3 t 2.3
seconds in 8/8 prospective trials, while the clinical team required 27.7 + 8.5 sec-
onds, and failed to initiate ictal SPECT in one of the trials.

10.2 Future Work

10.2.1 Feature Vector Enhancement

The feature vector developed in this thesis captures relationships between channels by
concatenating univariate spectral features derived from each EEG channel. Recently,
bivariate features that directly measure relationships between EEG channels, such as
phase synchrony, have been shown to be effective in seizure prediction [36]. More-
over, graph-theoretic features of a network representation of EEG, such as clustering
coefficient C (measure of local connectedness) and shortest path length L (measure
of overall network integration), have been shown to distinguish between the seizure
and non-seizure state [42]. Future work will investigate whether the addition of these
features to our feature vector improves the latency, sensitivity, or specificity of the
detector presented in Chapter 3.

10.2.2 Detecting Seizure Cessation

The detector presented in Chapter 3 signals the onset of a seizure, but is not designed
to signal the end of a seizure. Detecting the cessation of a seizure allows one to
compute seizure duration, and the length of the time interval between the end of
one seizure and the beginning of the next seizure. These quantities have clinical
significance. Seizures that persist for more than 10 minutes, or a cluster of seizures

that are closely spaced in time signal a possible transition into status epilepticus [13].

Status epilepticus refers to a life-threatening condition in which the brain enters a

state of persistent seizure activity. The detector in Chapter 3 learned the transition

155
between non-seizure and seizure activity using a feature vector that encodes the time
evolution of spectral and spatial features within the scalp EEG. Future work will
investigate whether the transition from seizure back to non-seizure activity can also
be learned using a similar feature vector.

10.2.3 Closed-Loop, Non-invasive Brain Stimulation

In this thesis we illustrated the feasibility of initiating vagus nerve stimulation follow-
ing the detection of seizure onset using non-invasive physiologic signals. Future work
will investigate the feasibility and therapeutic benefit of using our detection algo-
rithms to initiate other neurostimulation modalities such as as repetitive transcranial
magnetic stimulation (rTMS), transcranial direct current stimulation (tDCS), and
trigeminal nerve stimulation (TNS) [12]. This research will complement the develop-
ment of detect-and-treat systems that deliver cortical or deep-brain stimulation based
on the analysis of invasive physiologic signals (iEEG) [31, 41, 58].

156
Bibliography

[1] Rehan Akbani, Stephen Kwek, and Nathalie Japkowicz. Applying Support-Vector
Machines to Imbalanced Datasets. Springer, 2004.

[2] Al-Thaddeus Avestruz, Wesley Santa, Dave Carlson, Randy Jensen, Scott
Stanslaski, Alan Helfenstine, and Tim Denison. A 5yw/channel spectral analysis
ic for chronic bidirectional brain-machine interfaces. IEEE Journal Of Solid-State
Circuits, 43(12):3006-3024, 2008.

[3] E. Ben-Menachem. Vagus nerve stimulation for the treatment of epilepsy. Lancet
Neurology, 1(8):477-482, 2002.

[4] P. Boon, K. Vonck, and P. Van Walleghem et al. Programmed and magnet-
induced vagus nerve stimulation for refractory epilepsy. Journal Of Clinical
Neurophysiology, 18(5):402-407, 2001.

[5] Martin J Brodie, Steven C Schachter, and Patrick Kwan. Fast Facts: Epilepsy.
Health Press, third edition, 2005.

[6] Begley CE, Famulari M, Annegers JF, Lairson DR, Reynolds TF, and Coan S
et al. The cost of epilepsy in the United States: An estimate from population-
based and clinical survey. Epilepsia,41(3):342-351, 2000.

[7] Alexander M. Chan, Felice T. Sun, Erem H. Boto, and Brett M. Wingeier. Auto-
mated seizure onset detection for accurate onset time determination in intracra-
nial EEG. Clinical Neurophysiology, 119:2687-2696, 2008.

[8] Bernard S. Chang and Daniel H. Lowenstein. Epilepsy. The New England Journal
of Medicine, 349(13):1257-1266.

[9] Nello Cristianini and John Shawe-Taylor. An Introduction To Support Vector

Machines And Other Kernel Based Learning Methods. Cambridge University
Press, 2000.

[10] Barbara Cysyk, Jehuda Sepkuty, Ronald Lesser, and Ali Civelek. Truly ictal
spect is of major importance for reliable localization of seizure focus. Epilepsia,
38:Supplement 8:146, 1997.

157
[11] C.M. DeGiorgio, S.C. Shcahcter, A. Handforth, M. Salinsky, J. Thompson, and
B. Uthman et al. Prospective long-term study of vagus nerve stimulation for the
treatment of refractory seizures. Epilepsia, 41(9):1195-1200, 2000.

[12] C.M. DeGiorgio, A. Shewmon, D. Murray, and T. Whitehurst. Pilot study of

trigeminal nerve stimulation (TNS) for epilepsy: a proof-of-concept trial. Epilep-
sia, 47(7):1213-1215, 2006.

[13] Frank W. Drislane. Status Epilepticus. Humana Press, 2005.

[14] Javier Echauz, Stephen Wong, Otis Smart, Andrew Gardner, Gregory Worrell,
and Brian Litt. Computation applied to clinical epilepsy and antiepileptic de-
vices. In Ivan Soltesz and Kevin Staley, editors, ComputationalNeuroscience in
Epilepsy, part 32, pages 530-558. Academic Press, California, first edition, 2008.

[15] Michael Feichtinger, Hans Eder, Alexander Holl, Eva Korner, and Gerda Zmugg
et al. Automatic and remote controlled ictal SPECT injection for seizure focus
localization by use of a commercial contrast agent application pump. Epilepsia,
48(7):1409-1413.

[16] David M. Ficker. Sudden unexplained death and injury in epilepsy. Epilepsia,
41(Suppl. 2):S7-S12.

[17] David E. Friedman and Frank G. Gilliam. Seizure-related injuries are under-
rported in pharmacoresistant localization-related epilepsy. Epilepsia, 2009.

[18] J. Gotman. Automatic recognition of epileptic seizure in the EEG. Electroen-

cephalography and Clinical Neurophysiology, 54:530-540, 1982.

[19] J. Gotman, J.R. Ives, and P. Gloor. Frequency content of EEG and EMG at
seizure onset: Possibility of removal of EMG artifact by digital filtering. Elec-
troencephalographyand Clinical Neurophysiology, 52:626-639, 1981.

[20] Sukhi Grewal and Jean Gotman. An automatic warning system for epileptic
seizures recorded on intracerebral EEGs. Clinical Neurophysiology, 116:2460-
2472, 2005.

[21] The Vagus Nerve Stimulation Study Group. A randomized controlled trial of
chronic vagus nerve stimulation for treatment of medically intractable seizures.
Neurology, 45(2):224-230, 1995.

[22] E.J. Hammond, B.M. Uthman, S.A. Reid, and B.J. Wilder. Electrophysiological
studies of cervical vagus nerve stimulation in humans: EEG effects. Epilepsia,
33(6):1013-1020, 1992.

[23] A. Handforth, C.M. DeGiorgio, S.C. Schachter, B.M. Uthman, D.K. Naritoku,
and E.S. Tecoma et al. Vagus nerve stimulation therapy for partial onset seizures:
a randomized active control trial. Neurology, 51(1):48-55, 1998.

158
[24] C. Hopee, A. Poepel, and C.E. Elger. Epilepsy: Accuracy of patient seizure
counts. Arc Neurol, 64(11):1595-1599, 2007.

[25] G.L. Morris III. A retrospective analysis of the effects of magnet-activated stimu-
lation in conjuction with vagus nerve stimulation therapy. Epilepsy and Behavior,
4(6):740-745, 2003.

[26] J.D. Jirsch, E. Urrestarazu, P. LeVan, A. Olivier, F. Dubeau, and J. Gotman.

High-frequency oscillation during human focal seizures. Brain, 129:1593-1608,
2006.

[27] T. Joachims. Text Categorization with Support-Vector Machines: Learning with

Many Relevant Features. Springer, 1998.

[28] T. Joachims. Making large-Scale SVM Learning Practical.Advances in Kernel

Methods - Support Vector Learning. MIT-Press, 1999.

[29] Christophe Jouny and Gregory Bergey. Dynamics of epileptic seizures during
evolution and propagation. In Ivan Soltesz and Kevin Staley, editors, Com-
putational Neuroscience in Epilepsy, part 28, pages 457-469. Academic Press,
California, first edition, 2008.

[30] Eric Kandel, J.H. Schwartz, and Thomas M. Jessell. Principles Of Neural Sci-
ence. McGraw-Hill Medical, 2001.

[31] E.H. Kossoff, E.K. Ritzl, J.M. Politsky, A.M. Murro, J.R. Smith, and R.B. Duck-
row et al. Effect of an external responsive neurostimulator on seizures and electro-
graphic discharges during subdural electrode monitoring. Epilepsia,45(12):1560-
1567, 2004.

[32] D. Labar, J. Murphy, and E. Tecoma. Vagus nerve stimulation for medication-
resistant generalized epilepsy. Neurology, 52(7):1510-1512, 1999.

[33] Fritz Leutmezer, Christiana Schernthaner, Stefanie Lurger, Klaus Potzelberger,

and Christoph Baumgartner. Electrocardiographic changes at the onset of epilep-
tic seizures. Epilepsia, 44(3):348-354, 2003.

[34] P. LeVan, E. Urrestarzu, and J. Gotman. A system for automatic artifact re-
moval in ictal scalp EEG based on independent component analysis and bayesian
classification. ClinicalNeurophysiology, 117:912-927, 2006.

[35] Ralph Meier, Heike Dittrich, Andreas Schulze-Bonhage, and Ad Aertsen. Detect-
ing epileptic seizures in long-term human EEG: A new approach to automatic
online and reat-time detection and classification of polymorphic seizure patterns.
Journal of Clinical Neurophysiology, 25:119-131, 2008.

[36] Florian Mormann, Thomas Kreuz, Christoph Rieke, Ralph Andrzejak, Alexander
Kraskov, Peter David, Christian Elger, and Klaus Lehnertz. On the predictability
of epileptic seizures. Clinical Neurophysiology, 116:569-587, 2005.

159
[37] Anthony M. Murro, Don W. King, Joseph R. Smith, Brian B. Gallagher, Her-
man F. Flanigin, and Kimford Meador. Computerized seizure detection of com-
plex partial seizures. Clinical Neurophysiology, 79(4):330-333, 1991.

[38] Ivan Osorio, Mark Frei, and Steven Wilkinson. Real-time automated detection
and quantitative analysis of seizures and short term prediction of clinical onset.
Epilepsia, 39(6):615-627, 1998.

[39] Steven Pacia and John Ebersole. Intracranial EEG substrates of scalp ictal
patterns from temporal lobe foci. Epilepsia, 38(6):642-654, 1997.

[40] Shyamal Patel, Chiara Mancinelli, Anthony Dalton, Ben Patritti, Trudy Pang,
Steven Schachter, and Paolo Bonato. Detecting epileptic seizures using wearable
sensors. 35th Annual Northeast Bioengineering Conference, pages 1-2, 2009.

[41] T.E. Peters, N.C. Bhavaraju, M.G. Frei, and I. Osorio. Network system for au-
tomated seizure detection and contingent delivery of therapy. Journal of Clinical
Neurophysiology, 18(6):545-549, 2001.

[42] S.C. Ponten, F. Bartolomei, and C.J. Stam. Small-world networks and epilepsy:
Graph theoretical analysis of intracerebrally recorded mesial temporal lobe
seizures. Clinical Neurophysiology, 118:918-927, 2007.

[43] Hao Qu. Self-adapting Algorithms for Seizure Detection during EEG Monitering.
PhD dissertation, McGill University, 1995.

[44] Hao Qu and Jean Gotman. Improvement in seizure detection performance by

automatic adaptation to the EEG of each patient. Electroencephalographyand
Clinical Neurophysiology, 86:79-87, 1993.

[45] Hao Qu and Jean Gotman. A seizure warning system for long-term epilepsy
monitoring. Neurology, 45:2250-2254, 1995.

[46] Hao Qu and Jean Gotman. A patient-specific algorithm for the detection of
seizure onset in long-term EEG monitoring: Possible use as a warning device.
IEEE Transactions On Biomedical Engineering, 44:115-122, 1997.

[47] Steven Rothman and Xiao-Feng Yang. Local cooling: A therapy for intractable
neocortical epilepsy. Epilepsy Currents, 3(5):153-156, 2003.

[48] M.E. Saab and J. Gotman. A system to detect the onset of epileptic seizures in
scalp EEG. ClinicalNeurophysiology, 116:427-442, 2005.

[49] Steven Schachter. Brainstorms: Epilepsy in Our Words. Personal Accounts of

Living with Seizures. Raven Press, 1993.

[50] Steven Schachter and Clifford Saper. Vagus nerve stimulation. Epilepsia,39:677-
686, 1998.

160
[51] Steven C. Schachter. Vagus Nerve Stimulation. Informa HealthCare, 2003.

[52] B. Schlkopf, P. Knirsch, C. Smola, and A. Burges. Fast approximation of support

vector kernel expansions and an interpretation of clustering as approximation in
feature spaces. In P.Levi, M. Schanz, R.J. Ahler, and F.May, editors, Muster-
erkennung 1998-20 DA GM Symposium, pages 124-132. Springer-Verlag, Berline,
Germany, first edition, 1998.

[53] Jehuda P. Sepkuty, Ronal P. Lesser, Cahid, Barbara Cysyk, Robert Webber, and
Roy Shipley. An automated injection system (with patient selection) for SPECT
imaging in seizure localization. Epilepsia, 39(12):1350-1356.

[54] Eugene I. Shih, Ali H. Shoeb, and John V. Guttag. Sensor Selection for Energy-
Efficient Ambulatory Medical Monitoring. In Mobisys '09: Proceedings of the
7th internationalconference on Mobile systems, applications, and services, pages
347-358, New York, NY, USA, 2009. ACM.

[55] Ali Shoeb. Patient-Specific Seizure Onset Detection. MEng dissertation, Mas-
sachusetts Institute of Technology, 2003.

[56] Ali Shoeb, Herman Edwards, Jack Connolly, Blaise Bourgeois, S. Ted Treves, and
John Guttag. Patient-specific seizure onset detection. Epilepsy and Behavior,
5:483-498, 2004.

[57] Ali Shoeb, Trudy Pang, John Guttag, and Steven Schachter. Non-invasive com-
puterized system for automatically initiating vagus nerve stimulation following
patient-specific detection of seizures or epileptiform discharges. International
Journal of Neural Systems, 19(3):157-172, 2009.

[58] Felice Sun, Martha Morrell, and Robert Wharen. Responsive cortical stimulation
for the treatment of epilepsy. Neurotherapeutics,5(1):68-74, 2008.

[59] James Tao, Armit Ray, Susan Ebersle, and John Ebersole. Intracranial EEG
substrates of scalp eeg interictal spikes. Epilepsia, 46(5):669-676, 2005.

[60] William Theodore and Robert Fisher. Brain stimulation for epilepsy. Lancet
Neurology, 3(2):111-118, 2004.

[61] Sergios Theodoridis and Konstantinos Koutroumbas. PatternRecognition, chap-

ter 3. Elsevier, fourth edition, 2009.

[62] Elena Urrestarazu, Rahul Chander, Francois Dubeau, and Jean Gotman. Interic-
tal high-frequency oscillations (100-500hz) in the intracerebral EEG of epileptic
patients. Brain, 130:2354-2366, 2007.

[63] G.I. Varghese, M.J. Purcaro, and J.E. Motelaw et al. Clinical use of ictal SPECT
in secondarily generalized tonic-clonic seizures. Brain, 132:2102-2113.
[64] Scott Wilson, Mark Scheuer, Ronald Emerson, and Andrew Gabor. Seizure de-
tection: Evaluation of the Reveal algorithm. Clinical Neurophysiology, 10:2280-
2291, 2004.

[65] Scott B. Wilson. A neural network method for automatic and incremental learn-
ing applied to patient-dependant seizure detection. Clinical Neurophysiology,
116:1785-1795, 2005.

[66] Scott B. Wilson. Alogirthm architectures for patient dependant seizure detection.
Clinical Neurophysiology, 117:1204-1216, 2006.

[67] Dixon Woodbury and J. Walter Woodbury. Effects of vagal stimulation on ex-
perimentally induced seizures in rats. Epilepsia, 31(Suppl. 2):S7-S19, 1990.

[68] Jacob Zabara. Inhibition of experimental seizures in canines by repetitive vagal

stimulation. Epilepsia, 33(6):1005-1012, 1992.

[69] Feng Zhang, Li-Ping Wang, and Martin Brauner et al. Multimodal fast optical
interrogation of neural circuitry. Nature, 446:633-639, 2007.

162